Model Generalization in Deepfake Image Detection:

Datasets, Frequency Analysis, and Source Attribution

Executive Summary

The proliferation of deepfake technology necessitates robust and generalizable

detection mechanisms. This report addresses the critical aspects of achieving model
generalization for deepfake image detection, focusing on strategies for expanding
and diversifying datasets, leveraging frequency domain analysis to uncover subtle
forgery artifacts, and advancing methods for identifying the source of manipulated
content. The analysis highlights that deepfake detection is an ongoing "arms race,"
where generative models continuously evolve, producing increasingly sophisticated
forgeries that evade static detection approaches. Consequently, robust detection
systems must be inherently adaptive, capable of learning from diverse data,
identifying subtle frequency-domain anomalies, and attributing fakes to their
generative origins. Key recommendations include prioritizing datasets that encompass
diffusion models and "in-the-wild" characteristics, adopting spatial-frequency fusion
architectures, and employing ensemble and meta-learning techniques to enhance
cross-dataset performance. Practical considerations for Google Colab environments,
such as efficient data handling and cloud storage, are also discussed to facilitate
research and development.

1. Introduction to Deepfake Detection and the Generalization

Imperative

The landscape of digital media authenticity has been profoundly reshaped by the
rapid advancements in deepfake technology. These synthetic media, generated
through sophisticated artificial intelligence algorithms, pose significant challenges to
digital security and information integrity. The core objective of deepfake detection is
to reliably distinguish between genuine and manipulated content, a task complicated
by the continuous evolution of forgery techniques.

1.1. The Evolution of Deepfake Technology: From Autoencoders to Diffusion

Models

Deepfake technology has undergone several transformative stages, each contributing

to increased realism and complexity in synthetic media generation.

Initially, deepfake methods primarily relied on autoencoders.1 These models

compressed input data using an encoder and reconstructed them via a decoder, often
sharing a common encoder for tasks like face swapping. While demonstrating the
fundamental feasibility of face manipulation, the outputs from these early techniques
frequently exhibited smooth textures and lacked fine details, resulting in an artificial
appearance. These models generally required low-to-moderate GPU resources for
operation.1

The field experienced a significant leap with the introduction of Generative

Adversarial Networks (GANs) in 2014.1 GANs operate on an adversarial principle,
where a generator creates synthetic images and a discriminator simultaneously
evaluates their authenticity. This competitive process enables the generator to
produce highly realistic and high-resolution outputs, with prominent architectures
including StyleGAN and StarGAN.1 Despite these advancements, GAN-based
deepfakes often left discernible artifacts, such as mismatched lip movements,
inconsistencies in blinking, or variations in lighting, which detection systems
subsequently learned to exploit.1 Training GANs demands significantly higher
computational power compared to autoencoders due to their adversarial nature.1

Further sophistication led to the integration of synthetic voices with visual content,
resulting in highly convincing audio-visual deepfakes. Concurrently, neural textures
emerged, utilizing deferred neural rendering to push the boundaries of deepfake
realism.1 However, challenges persisted, including issues like mismatched lip
movements and visible splicing boundaries, particularly under complex lighting
conditions.1

The most recent innovation in deepfake generation is the adoption of diffusion

models.1 These models transform simple distributions, such as Gaussian noise, into
complex data that closely resemble real-world images or videos.1 While diffusion
models are highly advanced and capable of producing high-quality and diverse
outputs, they can still exhibit artifacts like residual noise and temporal inconsistencies
in video outputs.1 Training and inference with diffusion models generally require even
greater computational resources than GANs.1

The continuous progression from autoencoders to GANs and then to diffusion models
illustrates a dynamic often described as an "arms race" between deepfake creators
and their detectors.2 Early generative methods produced more overt, "low-order"
artifacts that were relatively straightforward to identify.1 As generative models became
more sophisticated, the nature of these detectable forgery traces evolved, shifting
from easily discernible flaws to increasingly subtle, high-frequency, or temporal
anomalies. For instance, diffusion models are noted for generating fewer detectable
artifacts and lacking the distinct "grid-like frequency artifacts" commonly associated
with GANs.3 This ongoing evolution means that any effective deepfake detection
system cannot rely on static, method-specific artifact identification. Instead, it must
be designed with inherent adaptability, perhaps through modular architectures or
continuous learning pipelines, focusing on fundamental, invariant traces of the
generative process rather than superficial cues. The core challenge is not merely
detecting current deepfakes but anticipating and countering future ones.

Furthermore, the escalating computational demands of deepfake generation—from

low-to-moderate for autoencoders to highest for diffusion models 1—suggests an
indirect relationship with detection. While greater computational power enables more
realistic fakes, it also implies that less sophisticated or resource-constrained
deepfake generation might leave more pronounced imperfections. This computational
signature embedded within a deepfake, reflecting the resources used for its creation,
could indirectly influence the type and prominence of artifacts. For example, a
deepfake produced quickly on consumer hardware might exhibit different, potentially
more obvious, imperfections compared to one generated with extensive cloud
computing resources. This factor could become relevant in advanced source
attribution, where understanding the generation environment might offer additional
clues.

1.2. The Critical Challenge of Model Generalization in Deepfake Detection

Despite significant advancements in deepfake detection, a persistent and critical
challenge remains: the poor generalization of models. Detectors often perform
exceptionally well on the datasets they were trained on but experience a substantial
decline in performance when confronted with unseen or out-of-distribution data.4

Several factors contribute to this generalization gap:

●​ Overfitting to Superficial Artifacts: Deepfake detectors frequently learn simple,
superficial "low-order interactions" among visual concepts. These interactions,
reflecting limited representations, can have a substantially negative impact on
deepfake detection as they do not generalize well to new forgeries.4 This
phenomenon means that models are overfitting to easily discernible, transient
artifacts rather than robust, fundamental forgery traces. When confronted with
novel, unseen forgeries, these biased representations fail, leading to significant
performance degradation during cross-dataset evaluations.4 To achieve effective
generalization, it is necessary to weaken or suppress these detrimental low-order
interactions, compelling models to learn more generalized artifact features.4 This
goes beyond simply providing more data; it requires innovative loss functions or
architectural designs that penalize such shallow learning.
●​ Domain Shift: A primary reason for the poor generalization of existing methods is
the significant feature distribution differences, or "domain shift," between labeled
source domain data (used for training) and large-scale unlabeled target domain
data (representing real-world, unseen deepfakes).7 Models trained on one
distribution struggle when applied to another, highlighting the need for methods
that can bridge this gap.
●​ Evolving Threat Landscape: The rapid advancement and increasing
sophistication of deepfake generation techniques ensure that detection models
are constantly playing catch-up.1 New forgery methods emerge quickly, rendering
older detection models less effective. This dynamic underscores that deepfake
detection is an ongoing "arms race".2 A detector that is state-of-the-art today
might be obsolete tomorrow. This necessitates a shift from building static, one-off
detection models to developing inherently adaptive and robust systems.
●​ Vulnerability to Adversarial Attacks: Deep learning-based deepfake detectors
are susceptible to adversarial examples—subtle, intentionally crafted
perturbations designed to deceive detection models.10 These attacks can
drastically reduce accuracy, further exacerbating generalization issues in
real-world, adversarial environments.

To address these challenges, several promising directions for improving generalization

have emerged:
 ●​ Data Augmentation and Transfer Learning: These techniques can enhance
generalization by synthetically expanding datasets to simulate a wider range of
forgery variations and by transferring knowledge from a source domain to a
target domain.7 However, these approaches often still require substantial amounts
of labeled data.7
●​ Meta-Learning and Domain Adaptation: Strategies such as the Open-World
Deepfake Detection Generalization Enhancement Training Strategy (OWG-DS)
aim to transfer deepfake detection knowledge from limited labeled data to
large-scale unlabeled data by reducing domain shift through feature distribution
alignment.7 The Multi-feature Channel Domain-Weighted (MCW) framework also
utilizes meta-learning, combining RGB and frequency domain information to
improve cross-database performance.17 These methods are crucial for learning
from limited labeled data and adapting to large-scale unlabeled target data.
●​ Ensemble Methods: Combining prediction probabilities from several
state-of-the-art asymmetric models has demonstrated more stable and reliable
performance across diverse datasets, mitigating the risk of a single model's
failure.9 This approach provides a layer of robustness by leveraging the collective
intelligence of multiple specialized detectors, which is particularly valuable in
dynamic threat landscapes where prior knowledge of forgery types is often
unavailable.
●​ Advanced Architectures: Research indicates that Transformer models
consistently outperform traditional Convolutional Neural Networks (CNNs) in
deepfake detection generalization.18 Models with multi-scale feature processing
capabilities, such as Res2Net-101, MViT-V2, and Swin Transformer, show superior
performance.18
●​ Self-Supervised Learning: Pre-training strategies like DINO have been shown to
yield superior feature representations and improve generalization compared to
purely supervised methods.18

The choice of training dataset significantly influences the generalizability of the

trained model. Datasets like FaceForensics++ and DFDC lead to comparably better
generalization capabilities compared to "easier" datasets, which tend to cause
overfitting.18 This highlights the critical importance of careful dataset selection and
expansion strategies to force models to learn more robust and less method-specific
features.

2. Expanding Datasets for Robust Generalization

The effectiveness of deepfake detection models is intrinsically linked to the quality,
diversity, and scale of the training data. For a project focused on model
generalization, strategic dataset expansion is not merely about increasing data
volume but about ensuring comprehensive representation of the evolving deepfake
landscape.

2.1. The Importance of Diverse and Large-Scale Datasets

Successfully training deep learning models for deepfake detection fundamentally

depends on access to sufficient, high-quality, and diverse data.19 Datasets must be
designed to emphasize generalization by including a wide variety of deepfake
generation techniques, such as autoencoders, GANs, and diffusion models, across
diverse contexts. This approach helps prevent models from overfitting to specific
artifacts.3

Real-world deepfakes are characterized by varying resolutions, compression levels,

diverse lighting conditions, and subtle temporal inconsistencies. Datasets should
ideally reflect this complexity to ensure that detectors are robust and perform reliably
in practical scenarios.3 Despite ongoing efforts, there remains a notable "lack of
large-scale ML-Generated Databases" that accurately represent the dynamic and
evolving landscape of deepfake creation, posing a significant challenge for consistent
performance evaluation across different research efforts.13

The necessity of large amounts of data for deep learning is well-established.19

However, the emphasis extends beyond mere quantity to the crucial role of

diversity in dataset design. Datasets must encompass a wide range of generation

methods and real-world contexts. For instance, studies have shown that "challenging"
datasets like FaceForensics++ (FF++) and the Deepfake Detection Challenge (DFDC)
lead to better generalization compared to "easier" ones.18 This suggests that simple
data volume is insufficient; the dataset must expose the model to a broad spectrum of
forgery artifacts and real-world variations, such as different compression levels. This
strategic diversity is vital for preventing models from overfitting to specific, transient
artifacts 24 and for achieving robust cross-dataset generalization.4
The intrinsic characteristics of a dataset directly influence a model's generalization
capabilities. Models trained on FF++ and DFDC consistently exhibit superior
generalization.18 This indicates that these datasets, by virtue of their comprehensive
inclusion of diverse manipulation methods and varying resolutions and compression,
compel models to learn more robust and less method-specific features. A more
challenging and diverse dataset, therefore, leads to models learning more invariant
and transferable features, resulting in improved cross-dataset performance. This
reinforces that dataset selection is a strategic decision that profoundly impacts the
ultimate utility of the detection model.

2.2. Recommended Publicly Available Datasets for Deepfake Detection

For a machine learning project focused on deepfake image detection and model
generalization, several publicly available datasets are highly recommended:
●​ FaceForensics++ (FF++): This is a foundational forensics dataset comprising
1000 original video sequences sourced from YouTube. These videos have been
manipulated using four automated face manipulation methods: Deepfakes,
Face2Face, FaceSwap, and NeuralTextures.6 FF++ also includes various
compression levels (raw/0, c23, c40), which are crucial for evaluating model
robustness to real-world data degradation.25 The raw video data is approximately
38.5GB, and raw extracted images can be as large as 2TB. However, compressed
versions (c23: ~10GB, c40: ~2GB) are significantly smaller and more
manageable.25 FF++ is a widely used benchmark, and models trained on it
(especially in combination with DFDC) have demonstrated comparably better
generalization capabilities.18 For Google Colab compatibility, using the
compressed versions is highly recommended due to storage and processing
limitations. Efficient loading strategies, such as​
ImageDataGenerator or mounting Google Drive, are essential.19
●​ Deepfake Detection Challenge (DFDC) Dataset: This is a large-scale dataset
designed specifically for deepfake detection, containing over 100,000 videos.6 It
offers a preview version with 5,000 videos and two facial modification algorithms,
and a full version with 124,000 videos featuring eight different facial modification
algorithms.22 The videos are sourced from paid actors, emphasizing data
diversity.22 Alongside FF++, DFDC is recognized for equipping models with
superior generalization capabilities.18 Given its large size, similar considerations as
FF++ apply for Google Colab, including the need for efficient loading and
 potentially a paid Colab plan for full dataset processing.19
●​ Deep Fake Detection (DFD) Entire Original Dataset: This dataset provides a
comprehensive collection of video sequences for deepfake detection,
downloaded from the official FaceForensics server. Its primary purpose is to
address the gap in available video-based original datasets for deepfake
detection.29 While the exact size is not specified in the available information, it is
implied to be of "substantial volume".29 It is suitable for training and evaluating
deep learning models for identifying manipulated media.29 As a video dataset, it
will likely require similar storage and efficient loading strategies as FF++ and
DFDC.19
●​ DFWild-Cup Competition Dataset: This dataset was provided for the 2025
Signal Processing Cup and contains a diverse range of deepfake images
generated from multiple deepfake image generators.20 Its explicit purpose is to
emphasize generalization in deepfake detection research.20 While its exact size is
not detailed, its design for a competition implies a substantial and challenging
collection of data. It is particularly valuable for projects focused on robust
generalization.
●​ StyleGAN-StyleGAN2 Deepfake Face Images Dataset: This dataset combines
and augments two existing datasets to ensure diversity, balance, and
generalizability.16 It includes 140,000 images (70,000 real from Nvidia's Flickr
dataset, 70,000 fake generated by StyleGAN and sourced from the DeepFake
Detection Challenge Discussion on Kaggle) and an additional 1,288 images (700
real from Unsplash, 588 fake from thispersondoesnotexist.com, utilizing
StyleGAN2).16 The dataset is further augmented with techniques like rotation,
shifting, brightness modification, zooming, cropping, and flipping, resulting in a
final size of 12,890 images.16 This dataset is explicitly curated for training and
evaluating deepfake detection models, studying GAN-generated images, and
exploring generalization techniques across multiple sources.16 Its relatively smaller
size makes it highly compatible with Google Colab for direct use or Google Drive
mounting.19
●​ Diffusion-generated Deepfake Detection dataset (D3): This large-scale
dataset addresses limitations in diversity and quantity found in older datasets. It
contains nearly 2.3 million records and 11.5 million images.21 Each record includes
a prompt, a real image, and four images generated by state-of-the-art
open-source diffusion models, specifically Stable Diffusion 1.4, 2.1, XL, and
DeepFloyd IF.21 The dataset also incorporates varying aspect ratios and
compression methods (BMP, GIF, JPEG, TIFF, PNG) to mimic real-world
distributions.21 D3 is crucial for addressing the detection of deepfakes generated
by diffusion models, which often lack the grid-like artifacts common in GANs.3
 Given its very large size, careful memory management, potentially using​
DiskDataset or iterating over samples, and mounting Google Drive for storage are
essential for Google Colab users.27
●​ DeepFakeFace (DFF) Dataset: This is a meticulously curated collection of
artificial celebrity faces generated using cutting-edge diffusion models, including
Stable Diffusion Inpainting, InsightFace, and Stable Diffusion V1.5.31 It also
includes original real images from the IMDB-WIKI dataset.31 The dataset is
distributed across four zip files, each containing 30,000 images, totaling 120,000
images.31 DFF is specifically designed to address challenges posed by deepfakes
generated by diffusion models.31 Its manageable size makes it suitable for Google
Colab, especially if mounted from Google Drive.27
●​ DF40: A Next-Generation Comprehensive Dataset: This dataset is a significant
advancement, comprising 40 distinct deepfake techniques, including recently
released State-of-the-Art (SOTA) methods.32 It is categorized into Face-swapping
(10 methods), Face-reenactment (13 methods), Entire Face Synthesis (12
methods, including StyleGAN2, StyleGAN3, StyleGAN-XL, SD-2.1, DDPM,
MidJourney6), and Face Edit (5 methods).32 DF40 provides both processed data
and original fake videos.32 Accepted by NeurIPS 2024, it is explicitly designed as a
"next-generation" dataset for comprehensive deepfake detection, making its
breadth of techniques exceptionally valuable for robust generalization research.32
Given its comprehensive nature, DF40 will likely be very large, necessitating
significant storage (Google Drive) and efficient processing strategies for Google
Colab users.32

The shift towards datasets that explicitly include deepfakes generated by diffusion
models is imperative. While foundational datasets like FF++ and DFDC are crucial,
newer datasets such as D3, DFF, and particularly DF40 directly address the evolving
threat. Diffusion models produce high-quality and diverse outputs that often do not
exhibit the noticeable grid-like artifacts in the frequency domain that older GANs do.3
This signifies a critical evolution in deepfake generation, meaning that detection
models trained solely on older GAN-based datasets will inherently struggle to
generalize to these newer, more sophisticated fakes. DF40, with its inclusion of 40
distinct techniques, including recent SOTAs from diffusion models, serves as an
excellent benchmark for training and evaluating next-generation deepfake detectors
that can handle the evolving threat landscape.32

While the project focuses on image detection, it is important to consider that

real-world deepfakes often exist within a broader, more complex multimedia context.
Datasets like "In-the-Wild" 33 are designed to evaluate a model's generalization to
realistic, uncontrolled samples. Even for image-based detection, understanding
temporal inconsistencies (common in video-based deepfakes) 6 or artifacts
introduced by video compression and streaming 17 can inform better feature extraction
for images. This suggests that while the current scope is images, researchers should
be mindful of how images are generated and disseminated within video contexts.
Datasets that explicitly capture "in-the-wild" characteristics beyond just diverse
generation methods, such as varying lighting, background complexities, and diverse
compression levels that mimic real-world distribution channels (as seen in D3's
inclusion of various compression methods) 21, can lead to more robust image
detectors.

For users operating in a Google Colab environment, strategic dataset management is

crucial. Datasets like raw FF++ images (~2TB) or the full DFDC (>100,000 videos) are
prohibitively large for direct use in Colab's free tier, which has resource limits and
ephemeral runtimes.25 The recommendation to use compressed versions (e.g., FF++
c23/c40) and mount Google Drive for storage is a direct consequence of these
limitations.23 Furthermore, employing

ImageDataGenerator for on-the-fly loading and augmentation 19 and understanding

DiskDataset for data larger than memory 30 are essential for efficient memory and I/O
management. Selecting datasets must involve a trade-off between diversity/size and
Colab's practical constraints. Adopting efficient data handling practices is necessary
to avoid runtime crashes and ensure a smooth development workflow. For very
large-scale experiments, a paid Colab plan or alternative computing resources might
eventually be necessary.

Table 1: Key Deepfake Datasets for Generalization Research

Dataset Type Key Size Deepfake Suitability Google

Name Characteri (approx.) Generatio for Colab
stics n Methods Generaliza Compatibi
Included tion lity Notes

FaceForen Video 1000 Raw: ~2TB Deepfakes High Use

sics++ original (images), , (widely compress
(FF++) videos, ~500GB Face2Face used ed
manipulat (videos); , benchmar versions
ed with 4 Compress FaceSwap, k, good (c23, c40).
methods, ed: ~10GB NeuralText for Mount
 various (c23), ures 26 cross-dat Google
compressi ~2GB aset Drive for
on levels. (c40) 25 generaliza storage.
tion when Use
combined ImageDat
with aGenerato
DFDC) 18 r for
efficient
loading. 19

Deepfake Video Over 100k Full: 124k Multiple (8 High Large

Detection videos, videos; algorithms (equips dataset,
Challenge diverse Preview: , not fully models consider
(DFDC) data from 5k videos specified with efficient
paid 22 but better loading
actors, 8 diverse) 28 generaliza and Colab
facial tion Pro/Enterp
modificati capabilitie rise for full
on s) 18 dataset.
algorithms Mount
. Google
Drive. 19

Deep Fake Video Comprehe Substantia Not Suitable Similar to

Detection nsive l volume explicitly for FF++ &
(DFD) collection (size not listed, but training/ev DFDC;
Entire of video specified) from aluating efficient
Original sequence 29 FaceForen deep loading
Dataset s from sics learning and
FaceForen server. models 29 storage
sics practices
server. needed. 19

DFWild-C Image Diverse Not Multiple Explicitly Manageab

up deepfake specified, deepfake designed le for
Competiti images but image for image-bas
on from designed generator generaliza ed, but
Dataset multiple for s 20 tion size will
generator competitio research determine
s. n. 20 20 feasibility.

StyleGAN- Image Combines 12,890 StyleGAN, Explicitly Highly

StyleGAN2 & images StyleGAN2 curated compatibl
Deepfake augments (final 16 for e due to
Face two augmente generaliza relatively
Images datasets, d) 16 tion smaller
includes across size.
real & sources 16 Mount
StyleGAN/ Google
StyleGAN2 Drive. 19
fakes.

Diffusion- Image 2.3M 11.5M Stable Crucial for Very large;

generated records, images 21 Diffusion addressin requires
Deepfake 11.5M 1.4, 2.1, XL, g careful
Detection images DeepFloyd diffusion- memory
(D3) from SOTA IF 21 generated managem
diffusion deepfakes ent
models, , designed (DiskDatas
varied for et),
aspect generaliza Google
ratios/com bility 21 Drive
pression. mounting.
27

DeepFake Image Curated 120,000 Stable Designed Manageab

Face artificial images (4 Diffusion to tackle le size for
(DFF) celebrity zip files, Inpainting, diffusion Colab,
Dataset faces from 30k each) InsightFac model especially
cutting-ed 31 e, Stable deepfakes with
ge Diffusion 31 Google
diffusion V1.5 31 Drive. 27
models.

DF40: Video/Ima 40 distinct Very large FSGAN, Exception Very large;

Next-Gen ge deepfake (not DeepFace ally processed
eration technique specified) Lab, valuable data more
Comprehe s, 32 FOMM, for robust manageab
nsive including SadTalker, generaliza le.
Dataset recent HeyGen, tion Requires
SOTAs StyleGAN2 research significant
(e.g., , due to storage
StyleGAN3 StyleGAN3 breadth of (Google
, SD-2.1, , technique Drive) and
HeyGen). StyleGAN- s 32 efficient
XL, SD-2.1, processin
DDPM, g. 32
MidJourne
y6,
StarGAN,
StyleCLIP,
 etc. 32

2.3. Strategies for Dataset Augmentation and Combination

To further enhance model generalization beyond the inherent diversity of selected

datasets, strategic data augmentation and combination techniques are indispensable.

Data augmentation is a vital technique for enhancing generalization by synthetically

expanding datasets and simulating a wider range of forgery variations and real-world
conditions.7 Common augmentation techniques include rotation, shifting, brightness
modification, zooming, cropping, and flipping.16 This process serves a purpose beyond
simply increasing the quantity of training data; it is a deliberate strategy to artificially
introduce

diversity that mimics the variations found in real-world deepfakes and their
dissemination, such as different lighting, angles, and compression artifacts. By
applying these transformations, the model is compelled to learn features that are
invariant to these common variations, directly improving its generalization to unseen,
real-world deepfakes. This is particularly crucial when obtaining truly diverse
"in-the-wild" data is challenging or impossible. Image augmentations have been
shown to be particularly helpful in improving the performance of Transformer models,
which are increasingly recognized for their strong generalization capabilities in
deepfake detection.18 Strategic and diverse data augmentation leads to models
learning more robust and invariant features, which in turn significantly improves their
generalization performance on out-of-distribution data.

Dataset combination involves merging and augmenting existing datasets from

distinct sources. For example, combining StyleGAN-generated fakes with real images
from different sources is a strategy to ensure diversity, balance, and generalizability,
preventing model overfitting to a single type of deepfake.16 This approach creates a
more comprehensive training environment, exposing the model to a broader spectrum
of forgery characteristics.

Ensemble learning is a powerful strategy to achieve robust cross-dataset

generalization by combining predictions from multiple state-of-the-art, asymmetric
models.9 The analysis indicates that no single model consistently outperforms others
across different settings, but ensemble-based predictions provide more stable and
reliable performance.9 This suggests that different models capture distinct types of
artifacts or generalize in varied ways. By combining their strengths, for example,
through probability averaging, a more resilient and generalizable detector can be
created. For a project explicitly focused on generalization, an ensemble approach is a
highly practical and recommended strategy. It mitigates the risk of a single model
overfitting or failing catastrophically on a novel deepfake type, providing a "top-down"
layer of robustness by leveraging the collective intelligence of multiple specialized
detectors. This approach is particularly valuable in dynamic threat landscapes where
prior knowledge of forgery types is often unavailable.

3. Frequency Domain Analysis for Enhanced Deepfake Detection

Deepfake detection has traditionally relied on analyzing visual cues in the spatial
domain. However, the increasing sophistication of generative models necessitates a
complementary approach that can uncover subtle, often imperceptible, forgery
traces. Frequency domain analysis offers precisely this advantage.

3.1. Why Frequency Domain? Uncovering Subtle Artifacts

While the spatial domain, which includes pixel values, colors, and textures, offers rich
feature representations for detecting image details and structures, most
spatial-based models tend to overfit to specific forgery characteristics.35 This
limitation often leads to poor generalization, especially when dealing with low-quality
or compressed images.35

Frequency domain analysis provides a crucial complementary perspective for

uncovering manipulation traces that are often imperceptible in the spatial domain.36
Deepfake manipulations frequently introduce subtle additive noise, unnatural spectral
distortion, or inconsistencies in high-frequency noise and upsampling operations.24
These anomalies manifest as distinctive spectral patterns that are more easily
detectable in the frequency domain. Techniques such as spectral analysis and phase
correlation are employed to reveal these hidden manipulations, including edge
discontinuities or subtle artifacts from generative models, which are difficult to
identify through spatial analysis alone.14 A purely spatial domain detector will
inherently miss these crucial forgery traces, particularly as deepfakes become more
visually convincing and subtle. Integrating frequency domain analysis is therefore
essential for identifying these hidden manipulations and significantly improving overall
detection accuracy and robustness 35, especially against compression artifacts.36

Furthermore, frequency domain features can enhance the resilience of detection

approaches against diverse image compression techniques, which often degrade
spatial features and make detection more challenging.36

However, a critical nuance in the application of frequency analysis is the evolving

nature of frequency artifacts across different generative models. While frequency
analysis is highly effective for detecting GAN-specific artifacts 24, diffusion models,
which represent the latest generation of deepfake technology, "do not exhibit
noticeable grid-like artifacts in the frequency domain".3 This means that the type of
frequency artifact changes with the generative model, posing a challenge for
traditional frequency-based methods designed for GANs. This necessitates a shift in
research towards identifying new types of frequency-domain artifacts introduced by
diffusion models, or developing more generalized frequency-aware techniques that
are not solely reliant on GAN-specific patterns. FreqNet, for example, attempts to
address this by learning "source-agnostic features" in the frequency domain,
suggesting a promising path forward.24

3.2. State-of-the-Art Techniques and Architectures

Current state-of-the-art deepfake detection models increasingly integrate spatial and

frequency domain analysis to achieve enhanced performance and generalization.

Spatial-Frequency Fusion Approaches:

●​ SFDT (Spatial-Frequency Domain Fusion Technology): This method
significantly improves deepfake video detection accuracy and model robustness
by fusing information from both spatial and frequency domains.36 It enhances
spatial feature extraction using improved Multiscale Vision Transformers (MVIT)
and Adaptive Structural Feature Fusion (ASFF) adaptive modules. Frequency
domain features are extracted via 2D Fast Fourier Transform (FFT), dynamic
filters, and the Frequency Enhancement Module (FECAM), which is designed to
reduce high-frequency noise.36 This air-frequency fusion scheme has
 demonstrated improved cross-compression detection performance and
robustness against various interferences.36
●​ SFCL (Spatial-Frequency Collaborative Learning) Framework: This framework
simultaneously extracts features from both spatial and frequency domains,
enabling multi-stage feature interaction through a Hierarchical Cross-Modal
Fusion (HCMF) mechanism.37 For frequency domain processing, SFCL departs
from conventional global frequency transformation and instead adopts a localized
analysis paradigm. Specifically, it employs block-wise Discrete Cosine Transform
(DCT) to precisely capture frequency domain artifacts while retaining macro-level
spatial semantics.37 This design aims for comprehensive representation and
generalizable detection.
●​ FMSI (Frequency-Domain Masking and Spatial Interaction) Model: This
innovative model introduces masked image modeling in frequency-domain
processing to prevent overfitting to specific frequency-domain features, thereby
enhancing generalization.35 It employs a dual-stream architecture for
spatial-frequency information interaction and leverages Vision Transformers to
capture global image features. The FMSI model also incorporates a
high-frequency information convolution module to prioritize high-frequency
processing, focusing on invariant features in the frequency domain to improve
generalization.35
●​ MCW (Multi-feature Channel Domain-Weighted) Framework: This framework
enhances generalizable deepfake detection, particularly in few-shot scenarios.17 It
improves a meta-learning network by combining RGB domain and frequency
domain information, and by assigning meta-weights to channels on the feature
map.17 This fusion and weighting strategy enhances feature extraction and the
model's ability to generalize by focusing on cross-domain invariant features.
●​ CFNSFD (Cross-Domain Fusion Network with One Spatial and Two
Frequency Domain) Features: This novel network architecture combines RGB
color features (spatial domain) with both shallow frequency features (extracted
via wavelet transformation) and deep frequency features (derived from residual
images using a specialized convolutional extractor).42 An adaptive feature fusion
module, utilizing gated convolution, integrates these multi-faceted features,
leading to high accuracy and robustness against diverse forgery techniques.42
This approach provides a more comprehensive description of image features,
enhancing robustness in detecting various forgery techniques.

FreqNet: Frequency-Aware Learning for Generalizability:

FreqNet is a novel, lightweight frequency-aware deepfake detection network specifically
designed to improve generalization across different GAN models, particularly when training
data is limited.24 Its core idea is to compel the detector to learn within the frequency domain,
rather than relying solely on source-specific frequency artifacts that typically arise from GAN
upsampling operations.24 The architecture of FreqNet includes a High-Frequency
Representation (HFR) module that forces the detector to prioritize high-frequency information
in both input images (via Fast Fourier Transform and high-pass filtering) and feature maps.39
A Frequency Convolutional Layer (FCL) further facilitates frequency space learning by
applying convolutions separately to the amplitude and phase spectra of feature maps,
allowing the network to learn source-agnostic features and reduce reliance on specific
training source characteristics.39 This approach aims to address the challenge posed by
diffusion models not exhibiting the same frequency artifacts as GANs, by enabling the model
to learn more generalized frequency-domain patterns.

4. Identifying the Source of Fake Images (Deepfake Attribution)

Beyond merely detecting whether an image is fake, the ability to attribute a deepfake
to its generative source is becoming increasingly important for forensic analysis,
intellectual property protection, and countering misinformation.

4.1. The Shift from Binary Classification to Source Attribution

Historically, research in digital image forensics primarily focused on binary

classification—determining whether an image was AI-generated or real.14 While this
approach provided a fundamental authenticity check, the rapid advancement of
deepfake technology and the emergence of new applications revealed its
insufficiency for regulating ethical use and ensuring accountability.45

This limitation has prompted a significant shift in research towards deepfake

attribution, which involves identifying the specific generative sources of
AI-generated images and providing evidence for this attribution.45 The goal is to move
beyond a simple fake/real label to pinpointing the generative model or even the
training dataset used to create the fake.44 This capability is crucial for holding
malicious actors accountable for AI-enabled disinformation, granting plausible
deniability to legitimate users of deepfake technology, and facilitating intellectual
property protection related to deepfake creation.45

The complexity and sheer number of available generative techniques, coupled with
the scarcity of high-quality, diverse open-source datasets specifically for this task,
make training and benchmarking synthetic image source attribution models highly
challenging.45 Current methods struggle with generalization, particularly in open-set
scenarios involving unknown or novel generators.45 Furthermore, deepfake detectors
are vulnerable to counter-forensic attacks, where malicious actors intentionally exploit
the limitations of existing detection methods.45

4.2. Methodologies for Attribution and Model Identification

Various methodologies are being developed to identify the training dataset or specific
generative model used to create deepfake images. These approaches often leverage
subtle, inherent "fingerprints" left by the generative process.

One novel forensic framework for identifying the training dataset (e.g., CelebA or
FFHQ) of GAN-generated images employs interpretable feature analysis.44 This
pipeline integrates spectral transforms, color distribution metrics, and local feature
descriptors to extract discriminative statistical signatures.44
●​ Spectral Transforms: Discrete Cosine Transform (DCT) and Fast Fourier
Transform (FFT) are applied to convert images to their spectral representations.
This allows for the identification of dataset-specific periodic artifacts and
compression-related signatures. For example, real datasets show smoother
frequency decay, while GAN-generated images often display localized
high-frequency artifacts or cross-shaped spectral artifacts, likely due to
convolutional upsampling patterns.38
●​ Color Distribution Metrics: Normalized RGB histograms capture color
information, revealing that real datasets maintain characteristic channel skews,
whereas GAN-generated images tend to smooth and normalize color profiles,
indicating implicit regularization strategies during GAN training.44
●​ Local Feature Descriptors: Scale-Invariant Feature Transform (SIFT) descriptors
are employed to model textural and structural properties.44​
​
These extracted features are then concatenated into a single feature vector and
standardized before being fed into supervised machine learning classifiers such
as Random Forest, Support Vector Machines (SVM), K-Nearest Neighbors (K-NN),
and XGBoost.38 This framework has achieved high accuracy (98-99%) in both
binary classification (real vs. synthetic) and multi-class dataset attribution across
 diverse GAN architectures.44 Frequency-domain features (DCT/FFT) are
particularly dominant in capturing dataset-specific artifacts.38

For identifying specific GAN architectures, the DNA-Det (Deepfake Network

Architecture Detector) method is a notable advancement.46 This method aims to
attribute fake images to their source GAN architectures even if the models have been
finetuned or retrained under different configurations (e.g., varying seeds, loss
functions, or datasets).46 DNA-Det is based on the observation that GAN architectures
tend to leave globally consistent fingerprints across an entire image, while traces left
by model weights can vary in different regions.46 Its methodology involves:
1.​ Pre-training on Image Transformation Classification: The network is
pre-trained to classify different image transformations (compression, blurring,
resampling, noising) on natural image datasets. This step helps the network learn
to focus on globally consistent traces related to fundamental image manipulation
operations, which are analogous to architecture-related traces, rather than
misleading semantic-related local features.46
2.​ Patchwise Contrastive Learning (PCL): PCL strengthens the global consistency
of extracted features by training on randomly cropped patches of images and
applying a contrastive loss. This encourages the network to learn features
consistent across different parts of the image, reinforcing the focus on
architecture-specific global fingerprints.46

By focusing on these globally consistent features, DNA-Det can isolate the unique
"fingerprints" left by the GAN's underlying architecture, leading to robust attribution
even in challenging cross-test setups (cross-seed, cross-loss, cross-finetune, and
cross-dataset).46

For diffusion models, which are increasingly prevalent, specific attribution challenges
arise because they often do not exhibit the same grid-like frequency artifacts as
GANs.3 However, research is progressing on identifying and attributing images from
text-to-image synthesis models like Stable Diffusion and DeepFloyd IF.48 Methods like
"DeepGuard" propose multi-classification techniques to attribute fake images to their
source models, achieving high accuracy.48 This involves training models to distinguish
between synthetic images created by various text-to-image models (e.g., Stable
Diffusion 1.5, SDXL, DeepFloyd IF) and authentic ones.48 The open-source nature of
models like Stable Diffusion allows for community involvement in developing detection
methods, focusing on inconsistencies like distorted fingers/hands, unnatural limbs,
blurry backgrounds, or garbled text, which are common artifacts.49
5. Practical Considerations for Google Colab Environment

Developing and experimenting with deepfake detection models, especially those

requiring large datasets and significant computational resources, presents unique
challenges within the Google Colab environment. Understanding and mitigating these
limitations is crucial for a smooth development workflow.

5.1. Handling Large Datasets in Colab

Google Colab, as a hosted Jupyter Notebook service, provides free access to

computing resources, including GPUs and TPUs, making it well-suited for machine
learning projects.34 However, its free tier comes with resource limits that can fluctuate,
and runtimes may terminate prematurely if not actively programmed in a notebook or
if abusive actions are detected, such as file hosting or cryptocurrency mining.34

For deepfake detection projects, which often involve substantial datasets, efficient
data handling is paramount. Datasets like FaceForensics++ can be extremely large
(e.g., ~2TB for raw extracted images, ~500GB for raw videos).25 Directly loading such
volumes into Colab's ephemeral runtime memory is often impractical.

The recommended strategy for handling large datasets in Colab is to load data
directly from Google Drive by using the mount drive method.23 This imports all
data from the user's Drive to the Colab runtime instance, providing persistent storage
beyond the session's ephemeral nature.23 For video datasets like FF++ and DFDC, it is
highly advisable to use

compressed versions (e.g., FF++ c23 or c40, which are significantly smaller at ~10GB
and ~2GB respectively) to reduce storage and processing demands.25

When working with image data, the ImageDataGenerator from Keras/TensorFlow can
be invaluable. Its flow_from_directory method helps read images from directories and
transform them into normalized NumPy arrays on the fly, reducing the need to load
the entire dataset into memory at once.19 This method allows specifying batch size
and target image size, which can be adjusted based on model requirements.19 For
datasets larger than memory, DeepChem's

DiskDataset class provides tools for efficiently working with data saved to disk,
ensuring it can be accessed even if the total amount exceeds available memory.30 This
approach involves iterating over the dataset, loading only small batches at a time,
processing them, and then freeing memory before loading the next batch.30

5.2. Optimizing Model Training and Evaluation in a Cloud Environment

Optimizing model training and evaluation in a Google Colab environment requires

careful attention to resource utilization and workflow efficiency.
●​ GPU/TPU Utilization: Colab provides free access to GPUs and TPUs, which are
essential for accelerating deep learning model training.34 Ensuring that the model
is configured to utilize these accelerators (e.g., by setting​
tf.distribute.MirroredStrategy for TensorFlow) is crucial. However, it is important to
note that Colab resources are not guaranteed and usage limits can fluctuate,
especially in the free tier.34 For demanding or long-running experiments,
purchasing a paid Colab plan (Colab Pro/Pro+) or exploring Google Cloud
Platform (GCP) Marketplace/Colab Enterprise for guaranteed resources might be
necessary.34
●​ Efficient Data Loading: As discussed, ImageDataGenerator and DiskDataset are
key for efficient data loading, preventing out-of-memory errors and speeding up
training by providing data in batches.19 Shuffling the training data using the
generator's​
shuffle=True option is important for robust model training, while setting
shuffle=False for validation/test sets ensures consistent evaluation order.19
●​ Callbacks for Training Stability: Implementing tf.keras.callbacks.EarlyStopping
is highly recommended.19 This callback monitors a specified metric (e.g.,​
val_loss) and stops training if it ceases to improve for a certain number of epochs
(patience), preventing overfitting and saving computational resources.19
●​ Model Architecture Selection: While larger, more complex models like
Transformer models often outperform CNNs in generalization 18, they also demand
more computational resources. For initial experimentation or
resource-constrained scenarios, lighter models like MobileNetV2 can be a good
starting point.19 MobileNetV2, for instance, offers a balance of acceptable
accuracy and lower computational demands, making it suitable for environments
 with limited processing resources.6
●​ Evaluation Metrics: Beyond simple accuracy, a comprehensive evaluation should
include metrics like confusion matrices and classification reports (precision,
recall, F1-score).19 For generalization, Area Under the Receiver Operating
Characteristic curve (AUROC) and Area Under the Precision-Recall Curve
(AUPRC) are particularly important, as they assess performance across various
thresholds and are less sensitive to class imbalance.9
●​ Version Control and Reproducibility: While not explicitly a Colab feature, using
GitHub for version control and sharing notebooks is critical for reproducibility.
Colab notebooks can be directly linked to GitHub, allowing for easy collaboration
and tracking of changes.23 This aligns with the broader research community's
need for reproducible results, including detailed experimental setups and
open-source tools.13

6. Conclusions and Recommendations

The challenge of deepfake image detection, particularly concerning model

generalization, is a complex and evolving problem. The analysis presented in this
report underscores that effective solutions must move beyond static detection
methods to embrace dynamic, adaptive, and comprehensive approaches. The
continuous "arms race" between deepfake generation and detection necessitates a
proactive stance, where detection systems are designed to anticipate and counter
future, more sophisticated forgeries.

Key Conclusions:
●​ Dynamic Threat Landscape: Deepfake technology is rapidly advancing, with
diffusion models introducing new types of artifacts that differ from those
generated by older GANs. This means that detection models must constantly
adapt, and reliance on methods that overfit to "toxic" low-order interactions or
specific, transient artifacts will lead to poor generalization.
●​ Dataset Diversity is Paramount: The sheer volume of data is not enough.
Generalizable models require training on diverse datasets that encompass a wide
range of deepfake generation techniques (including diffusion models), varying
compression levels, and "in-the-wild" characteristics. Datasets like FF++, DFDC,
D3, DFF, and particularly DF40 are crucial for this purpose.
●​ Frequency Domain Analysis is Essential: Spatial-domain analysis alone is
 insufficient. Frequency domain analysis offers a complementary lens to uncover
subtle, often imperceptible, forgery traces and enhance robustness against image
compression. The development of frequency-aware architectures and
spatial-frequency fusion techniques is critical.
●​ Source Attribution is a Growing Necessity: Moving beyond binary
classification, the ability to identify the specific generative model or training
dataset of a deepfake is vital for forensic analysis, accountability, and intellectual
property protection. This requires specialized attribution methodologies that can
detect architectural fingerprints.
●​ Practicality in Cloud Environments: For researchers utilizing platforms like
Google Colab, efficient data handling strategies (e.g., Google Drive mounting,
ImageDataGenerator, DiskDataset) are not merely optimizations but necessities
for managing large datasets and ensuring productive research.

Actionable Recommendations:
1.​ Prioritize Diverse Dataset Acquisition and Generation:
○​ Actively seek and integrate datasets that include deepfakes generated by
diffusion models (e.g., D3, DFF, DF40) to prepare models for the latest
generation of forgeries.
○​ Emphasize datasets with "in-the-wild" characteristics, including varying
resolutions, compression levels, and diverse environmental factors, to
enhance real-world applicability.
○​ Consider synthetic data generation and strategic data augmentation
(e.g., rotation, shifting, brightness, cropping, flipping) to further expand
dataset diversity and simulate unseen variations, especially for training
Transformer models.
2.​ Adopt Spatial-Frequency Fusion Architectures:
○​ Implement and experiment with hybrid detection frameworks that integrate
both spatial and frequency domain analysis (e.g., SFDT, SFCL, FMSI, CFNSFD).
This dual-domain approach is crucial for capturing both overt visual
inconsistencies and subtle spectral artifacts.
○​ Explore frequency-aware learning techniques like FreqNet, which compel
models to learn generalized, source-agnostic features in the frequency
domain, rather than overfitting to specific GAN-induced grid patterns.
3.​ Implement Generalization-Enhancing Training Strategies:
○​ Utilize ensemble learning approaches by combining predictions from
multiple asymmetric models. This strategy provides more stable and reliable
performance across diverse, unseen datasets, mitigating the risk of individual
model failures.
 ○​ Investigate meta-learning and domain adaptation techniques (e.g.,
OWG-DS, MCW) to enable models to learn from limited labeled data and
adapt effectively to large-scale unlabeled target domains, addressing the
critical domain shift problem.
○​ Leverage self-supervised pre-training strategies (e.g., DINO) for backbone
models to learn superior feature representations that generalize better across
different deepfake types.
4.​ Develop and Integrate Deepfake Attribution Capabilities:
○​ Incorporate methodologies for source dataset attribution (e.g., using
spectral transforms, color distribution metrics, and local feature descriptors)
to identify the training data used for GAN-generated images.
○​ Explore techniques for generative model architecture attribution (e.g.,
DNA-Det) to pinpoint the specific deepfake generation tool, even if it has
been finetuned or retrained. This is crucial for forensic analysis and
accountability.
5.​ Optimize for Google Colab Environment:
○​ Always mount Google Drive for persistent storage of large datasets and
model checkpoints.
○​ Prioritize using compressed versions of large video datasets (e.g., FF++
c23/c40) to manage resource consumption.
○​ Employ efficient data loading mechanisms like ImageDataGenerator for
image data and consider DiskDataset for datasets exceeding memory
capacity.
○​ Utilize callbacks (e.g., EarlyStopping) during training to prevent overfitting
and optimize computational resource usage.
○​ Be prepared to consider Colab Pro/Pro+ for more demanding experiments
that require guaranteed resources and longer runtimes.

By systematically addressing these recommendations, the deepfake detection project

can significantly enhance its models' generalization capabilities, enabling them to
effectively counter the ever-evolving threat of synthetic media.

Works cited

1.​ Generative Artificial Intelligence and the Evolving Challenge of ..., accessed July
19, 2025, https://www.mdpi.com/2224-2708/14/1/17
2.​ The Evolution of Deepfake Technology: From Simple Face Swaps to AI-Powered
Illusions | by SingularityNET Ambassadors - Medium, accessed July 19, 2025,
https://medium.com/singularitynet-ambassador-community-page/the-evolution-
of-deepfake-technology-from-simple-face-swaps-to-ai-powered-illusions-59fe
88d89b96
3.​ Diffusion Deepfake - OpenReview, accessed July 19, 2025,
https://openreview.net/pdf/add2d2cf12b751ad53cb1501fa1937a9b42db92b.pdf
4.​ Towards Understanding the Generalization of ... - CVF Open Access, accessed
July 19, 2025,
https://openaccess.thecvf.com/content/ICCV2023/papers/Yao_Towards_Understa
nding_the_Generalization_of_Deepfake_Detectors_from_a_Game-Theoretical_IC
CV_2023_paper.pdf
5.​ Towards the Detection of Diffusion Model Deepfakes - SciTePress, accessed July
19, 2025, https://www.scitepress.org/Papers/2024/124220/124220.pdf
6.​ A SURVEY ON: DEEPFAKE DETECTION SYSTEM - IRJMETS, accessed July 19, 2025,
https://www.irjmets.com/uploadedfiles/paper//issue_11_november_2024/63699/fin
al/fin_irjmets1731397529.pdf
7.​ Towards Open-world Generalized Deepfake Detection: General Feature
Extraction via Unsupervised Domain Adaptation - arXiv, accessed July 19, 2025,
https://arxiv.org/html/2505.12339v1
8.​ www.jas.shu.edu.cn, accessed July 19, 2025,
https://www.jas.shu.edu.cn/EN/10.3969/j.issn.0255-8297.2025.03.007
9.​ Ensemble-Based Deepfake Detection using State-of-the-Art ... - arXiv, accessed
July 19, 2025, https://arxiv.org/abs/2507.05996
10.​Comprehensive Evaluation of Deepfake Detection Models: Accuracy ..., accessed
July 19, 2025, https://www.mdpi.com/2076-3417/15/3/1225
11.​ A Comprehensive Survey of DeepFake Generation and Detection ..., accessed
July 19, 2025, https://www.icck.org/article/abs/jiap.2025.431672
12.​A Comprehensive Evaluation of Deepfake Detection Methods: Approaches,
Challenges and Future Prospects | ITM Web of Conferences, accessed July 19,
2025,
https://www.itm-conferences.org/articles/itmconf/abs/2025/04/itmconf_iwadi202
4_03002/itmconf_iwadi2024_03002.html
13.​Deepfakes Generation and Detection: A Short Survey - PMC, accessed July 19,
2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC9863015/
14.​A Survey: Deepfake and Current Technologies for ... - CEUR-WS.org, accessed
July 19, 2025, https://ceur-ws.org/Vol-3900/Paper9.pdf
15.​Adaptive DCGAN and VGG16-based transfer learning for effective deepfake
detection - IET Digital Library, accessed July 19, 2025,
https://digital-library.theiet.org/doi/pdf/10.1049/icp.2025.1711?download=true
16.​StyleGan-StyleGan2 Deepfake Face Images - Kaggle, accessed July 19, 2025,
https://www.kaggle.com/datasets/kshitizbhargava/deepfake-face-images
17.​MCW: A Generalizable Deepfake Detection Method for Few-Shot ..., accessed
July 19, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC10649340/
18.​(PDF) Deepfake Detection: Analysing Model Generalisation Across ..., accessed
July 19, 2025,
https://www.researchgate.net/publication/376975669_Deepfake_Detection_Analy
sing_Model_Generalisation_Across_Architectures_Datasets_and_Pre-Training_Pa
radigms
19.​DeepFake Detection.ipynb - Colab, accessed July 19, 2025,
 https://colab.research.google.com/drive/1GRyxgtxpxGNLQkvcUcpF3cXEnFQSNpI
3?usp=sharing
20.​Hybrid Deepfake Image Detection: A Comprehensive Dataset-Driven Approach
Integrating Convolutional and Attention Mechanisms with Frequency Domain
Features - arXiv, accessed July 19, 2025, https://arxiv.org/html/2502.10682v1
21.​Diffusion-generated Deepfake Detection dataset (D 3 ) - AImageLab - Unimore,
accessed July 19, 2025,
https://aimagelab.ing.unimore.it/imagelab/go.asp?string=d3-dataset
22.​[1910.08854] The Deepfake Detection Challenge (DFDC) Preview Dataset - ar5iv -
arXiv, accessed July 19, 2025, https://ar5iv.labs.arxiv.org/html/1910.08854
23.​Deepfakes with the First Order Model Method - Colab - Google, accessed July
19, 2025,
https://colab.research.google.com/github/JaumeClave/deepfakes_first_order_mo
del/blob/master/first_order_model_deepfakes.ipynb
24.​Frequency-Aware Deepfake Detection: Improving Generalizability through
Frequency Space Learning - arXiv, accessed July 19, 2025,
https://arxiv.org/html/2403.07240v1
25.​FaceForensics/dataset/README.md at master - GitHub, accessed July 19, 2025,
https://github.com/ondyari/FaceForensics/blob/master/dataset/README.md
26.​FaceForensics++ Dataset - Papers With Code, accessed July 19, 2025,
https://paperswithcode.com/dataset/faceforensics-1
27.​neptune.ai, accessed July 19, 2025,
https://neptune.ai/blog/how-to-use-google-colab-for-deep-learning-complete-t
utorial#:~:text=To%20train%20complex%20models%2C%20you,where%20the%
20dataset%20is%20stored.
28.​DFDC Dataset - Papers With Code, accessed July 19, 2025,
https://paperswithcode.com/dataset/dfdc
29.​Deep Fake Detection (DFD) Entire Original Dataset - Kaggle, accessed July 19,
2025,
https://www.kaggle.com/datasets/sanikatiwarekar/deep-fake-detection-dfd-entir
e-original-dataset
30.​Working With Datasets - Colab, accessed July 19, 2025,
https://colab.research.google.com/github/deepchem/deepchem/blob/master/exa
mples/tutorials/Working_With_Datasets.ipynb
31.​OpenRL/DeepFakeFace · Datasets at Hugging Face, accessed July 19, 2025,
https://huggingface.co/datasets/OpenRL/DeepFakeFace
32.​YZY-stack/DF40: Official repository for the next-generation ... - GitHub, accessed
July 19, 2025, https://github.com/YZY-stack/DF40
33.​In-the-Wild Dataset - Deepfake Total, accessed July 19, 2025,
https://deepfake-total.com/in_the_wild
34.​Google Colab, accessed July 19, 2025,
https://research.google.com/colaboratory/faq.html
35.​Frequency-Domain Masking and Spatial Interaction for Generalizable Deepfake
Detection, accessed July 19, 2025, https://www.mdpi.com/2079-9292/14/7/1302
36.​(PDF) Deepfake Detection Technology Integrating Spatial Domain and Frequency
 Domain, accessed July 19, 2025,
https://www.researchgate.net/publication/390254350_Deepfake_Detection_Tech
nology_Integrating_Spatial_Domain_and_Frequency_Domain
37.​Towards Generalizable Deepfake Detection with Spatial-Frequency Collaborative
Learning and Hierarchical Cross-Modal Fusion - arXiv, accessed July 19, 2025,
https://arxiv.org/html/2504.17223v1
38.​Deepfake Forensic Analysis: Source Dataset Attribution and Legal Implications of
Synthetic Media Manipulation - arXiv, accessed July 19, 2025,
https://www.arxiv.org/pdf/2505.11110
39.​[Papierüberprüfung] Frequency-Aware Deepfake Detection: Improving
Generalizability through Frequency Space Learning - Moonlight | AI Colleague for
Research Papers, accessed July 19, 2025,
https://www.themoonlight.io/de/review/frequency-aware-deepfake-detection-im
proving-generalizability-through-frequency-space-learning
40.​[Literature Review] Frequency-Aware Deepfake Detection: Improving
Generalizability through Frequency Space Learning - Moonlight, accessed July 19,
2025,
https://www.themoonlight.io/review/frequency-aware-deepfake-detection-impro
ving-generalizability-through-frequency-space-learning
41.​Deepfake Detection Technology Integrating Spatial Domain and Frequency
Domain | Frontiers in Computing and Intelligent Systems - Darcy & Roy Press,
accessed July 19, 2025, https://drpress.org/ojs/index.php/fcis/article/view/30019
42.​(PDF) Deepfake Detection Based on the Adaptive Fusion of Spatial ..., accessed
July 19, 2025,
https://www.researchgate.net/publication/385649528_Deepfake_Detection_Base
d_on_the_Adaptive_Fusion_of_Spatial-Frequency_Features
43.​arxiv.org, accessed July 19, 2025, https://arxiv.org/pdf/2403.07240v1
44.​(PDF) Deepfake Forensic Analysis: Source Dataset Attribution and ..., accessed
July 19, 2025,
https://www.researchgate.net/publication/391856499_Deepfake_Forensic_Analysi
s_Source_Dataset_Attribution_and_Legal_Implications_of_Synthetic_Media_Mani
pulation
45.​Deepfake attribution: On the source identification of artificially ..., accessed July
19, 2025,
https://www.researchgate.net/publication/356770322_Deepfake_attribution_On_t
he_source_identification_of_artificially_generated_images
46.​Deepfake Network Architecture Attribution - Association for the ..., accessed July
19, 2025, https://cdn.aaai.org/ojs/20391/20391-13-24404-1-2-20220628.pdf
47.​Deepfake Network Architecture Attribution | Proceedings of the AAAI
Conference on Artificial Intelligence, accessed July 19, 2025,
https://ojs.aaai.org/index.php/AAAI/article/view/20391
48.​DeepGuard: Identification and Attribution of AI-Generated Synthetic ..., accessed
July 19, 2025,
https://researchportal.port.ac.uk/files/100827541/electronics-14-00665.pdf
49.​Stable Diffusion Deepfakes: Creation and Detection | by Tahir - Medium, accessed
 July 19, 2025,
https://medium.com/@tahirbalarabe2/stable-diffusion-deepfakes-creation-and-d
etection-15103f99f55d
50.​deep-floyd/IF - GitHub, accessed July 19, 2025, https://github.com/deep-floyd/IF
51.​Stability AI releases DeepFloyd IF, a powerful text-to-image model that can
smartly integrate text into images, accessed July 19, 2025,
https://stability.ai/news/deepfloyd-if-text-to-image-model