PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepFakes and Beyond: A Survey of Face Manipulation and Fake Detection

Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales, Javier Ortega-Garcia

arXiv:2001.00179v3cs.CVcs.MM

TL;DR

Realistic facial manipulations create concerns including fake news, hoaxes, and fraud, motivating improved detection methods. This survey reviews four major manipulation types, their databases and detection benchmarks, with special attention to DeepFakes and detection challenges.

  • Problem

    Realistic manipulated facial content can support fake news, hoaxes, and financial fraud, while existing detection methods require improvement against increasingly sophisticated content.

  • Method

    The survey reviews four facial manipulation types, associated techniques and public databases, and benchmarks for evaluating fake detection methods, emphasizing recent DeepFakes.

  • Results

    The reviewed benchmarks generally show low detection error rates under controlled conditions, but detection performance degrades under unseen scenarios and post-processing conditions.

  • Takeaways & Limitations

    Fairer comparisons and stronger detector generalisation require attention to differing databases, compression levels, metrics, protocols, and unseen conditions.

  • Takeaways & Limitations

    Most surveyed detectors are evaluated under controlled conditions, whereas shared social-media content varies in compression, resizing, and noise.

Abstract

from arXiv · show

The free access to large-scale public databases, together with the fast progress of deep learning techniques, in particular Generative Adversarial Networks, have led to the generation of very realistic fake content with its corresponding implications towards society in this era of fake news. This survey provides a thorough review of techniques for manipulating face images including DeepFake methods, and methods to detect such manipulations. In particular, four types of facial manipulation are reviewed: i) entire face synthesis, ii) identity swap (DeepFakes), iii) attribute manipulation, and iv) expression swap. For each manipulation group, we provide details regarding manipulation techniques, existing public databases, and key benchmarks for technology evaluation of fake detection methods, including a summary of results from those evaluations. Among all the aspects discussed in the survey, we pay special attention to the latest generation of DeepFakes, highlighting its improvements and challenges for fake detection. In addition to the survey information, we also discuss open issues and future trends that should be considered to advance in the field.

I. INTRODUCTION

The survey frames increasingly realistic facial manipulations as a media-forensics concern and organizes them into four principal groups. It reviews manipulation methods, databases, detection benchmarks, and related open issues.

  • Motivation: DeepFake content can support fake news, hoaxes, financial fraud, and other harmful applications.
  • Scope: The survey reviews entire face synthesis, identity swap, attribute manipulation, and expression swap, while also discussing face morphing.
  • Scope: The survey complements earlier work with detailed reviews of manipulation techniques, public databases, detection benchmarks, and benchmark results for each manipulation group.
  • Manipulation groups: Entire face synthesis generates non-existent faces, identity swap replaces one person’s face with another’s, and attribute manipulation changes facial properties such as age or hair color.
  • Manipulation groups: Expression swap modifies a person’s facial expression, with Face2Face and NeuralTextures highlighted as video techniques.

III. ENTIRE FACE SYNTHESIS

Entire face synthesis produces realistic non-existent faces, and public databases support detection research across ProGAN and StyleGAN imagery. GANprintR makes iFakeFaceDB especially challenging by removing GAN fingerprints while preserving realistic appearance.

  • Manipulation and databases: Entire face synthesis creates non-existent face images, commonly using GAN architectures such as ProGAN and StyleGAN.
  • Manipulation and databases: Detection experiments require real face images from separate public databases because the reviewed synthesis databases contain fake images generated by GAN architectures.
  • Manipulation and databases: StyleGAN learns an unsupervised separation between high-level attributes and stochastic variation, enabling scale-specific control of face synthesis.
  • Public databases: 100K-Faces contains 100,000 StyleGAN images trained on controlled-background photos, excluding strange StyleGAN background artifacts.
  • Public databases: 250,000 StyleGAN images and 80,000 ProGAN images comprise iFakeFaceDB, whose GANprintR processing removes GAN fingerprints and increases detection difficulty.

B. Manipulation Detection

The survey compares detection approaches for synthetic face images, covering feature-based, neuron-activation, convolutional-trace, steganalysis, fingerprint, and incremental-learning methods. Results are strong in matched settings but degrade under dataset mismatch, unseen scenarios, perturbations, and challenging databases.

  • Detection approaches: 70.0% AUC was achieved using colour features and a linear SVM on the NIST MFC2018 dataset.The approach analyzes colour differences between real camera images and fake synthesis images.
  • Detection approaches: 84.7% fake detection accuracy was achieved by FakeSpoter using neuron-coverage features from FaceNet.FakeSpoter extracts neuron-coverage behaviors from deep face-recognition systems and classifies them with an SVM.
  • Detection approaches: 99.81% Acc. was achieved by analyzing convolutional traces extracted with Expectation Maximization and classified using k-NN, SVM, or LDA.The evaluation included fakes generated with AttGAN, GDWCT, StarGAN, StyleGAN, and StyleGAN2.
  • Detection approaches: 12.3% EER was reported for a steganalysis-inspired detector evaluated on the 100K-Faces database.The detector combines pixel co-occurrence matrices with CNNs and was also examined for cross-GAN generalization.
  • Detection approaches: 99.5% fake detection accuracy was achieved by a GAN-fingerprint attribution network, but robustness to unseen noise, blur, cropping, and compression was limited without retraining.The method maps images to fingerprint representations and classifies their correlation with learned source fingerprints.
  • Generalization and evaluation: Detection performance deteriorated under dataset mismatch and unseen scenarios, with iFakeFaceDB producing a best EER of 4.5%.Controlled evaluations reached EER = 0.02%, whereas cross-source and cross-data conditions exposed substantial generalization difficulty.

IV. IDENTITY SWAP

Identity swap replaces one person’s face with another’s in realistic fake videos, and its public databases have evolved from controlled, artifact-prone first-generation data to more realistic and variable second-generation data. The section reviews these databases and summarizes the improvements in visual quality and in-the-wild diversity.

  • Manipulation technique: Identity swap replaces one person’s face with another’s to generate realistic fake videos rather than manipulated still images.This distinguishes identity swap from entire face synthesis, which operates at image level.
  • First-generation databases: First-generation databases include UADFV, DeepfakeTIMIT, and FaceForensics++, with early datasets often using limited identities, controlled settings, and visible artifacts.UADFV, for example, contains 49 real videos transformed into 49 fakes using Nicolas Cage as the swapped identity.
  • Second-generation databases: Second-generation databases include Celeb-DF and DFDC, developed to provide higher visual quality and more realistic Internet-like or diverse video content.Celeb-DF targets improved visual quality, while DFDC was released as a large-scale challenge dataset.
  • Generation differences: First-generation fakes commonly show low-quality synthesis, colour mismatch, mask boundaries, retained facial elements, limited pose variation, and temporal artifacts.These databases also generally use controlled camera-position and lighting conditions.
  • Generation differences: Second-generation databases improve visual realism and variability through in-the-wild conditions, including varied environments, lighting, camera distances, and poses.DFDC specifically includes indoor and outdoor acquisition, day and night lighting, varying subject-to-camera distances, and pose variations.

B. Manipulation Detection

Identity-swap detection uses image, temporal, facial-motion, and attention-based cues, with strong benchmark results but weaker generalization on newer databases.

  • Identity-swap detection: Facial expressions, head movements, blinking, and eye-related features provide alternative detection cues, including 87.5% accuracy on a proprietary database.These approaches target inconsistencies in facial motion and physiological or geometric signals.
  • Identity-swap detection: XceptionNet generally provided the best results for DeepFakes and FaceSwap across evaluated video-quality levels.The survey notes that processing quality affects detection performance.
  • Identity-swap detection: Capsule Networks exceeded 90% accuracy on FaceForensics++ but showed poor generalization on unseen databases.The same pretrained model was tested against databases not used in its original evaluation.
  • Identity-swap detection: 99.43% AUC and 3.1% EER were achieved with an attention-based detector on DFFD identity-swap data.The authors describe the approach as producing state-of-the-art results, while noting protocol differences complicate comparisons.
  • Identity-swap detection: Temporal-aware CNN-RNN and recurrent convolutional methods exploit discrepancies across frames to detect fake videos.The recurrent convolutional approach achieved 96.9% and 96.3% AUC for reported DeepFake settings.
  • Identity-swap detection: Detection performance declined on second-generation DeepFake databases, reaching 91.0% AUC on DFDC Preview and 83.6% on Celeb-DF.The comparison was reported relative to first-generation databases.

A. Manipulation Techniques and Public Databases

Attribute manipulation research covers GAN-based editing methods and comparatively limited public data, while detection benchmarks compile results across several approaches.

  • Manipulation techniques and databases: Attribute manipulation changes facial properties such as hair or skin colour, gender, age, or the presence of glasses.The survey frames these edits as image-level facial-attribute changes.
  • Manipulation techniques and databases: Public availability of GAN code lets researchers generate custom datasets, contributing to the limited number of publicly available attribute-manipulation databases.The survey therefore emphasizes reviewing GAN approaches alongside available research resources.
  • Manipulation techniques and databases: StarGAN performs image-to-image translation across multiple domains using a single model rather than separate models for every domain pair.This design addresses the difficulty of handling more than two domains directly.
  • Manipulation techniques and databases: AttGAN applies attribute classification to generated images and reports realistic manipulation while preserving other facial details.It removes the strict attribute-independent constraint from the latent representation.
  • Manipulation techniques and databases: STGAN incorporates selective transfer units to improve attribute manipulation and image quality beyond encoder-decoder bottlenecks.The motivation is that bottleneck layers can produce blurry, low-quality results.
  • Manipulation techniques and databases: The DFFD database contains 18,416 FaceApp-generated and 79,960 StarGAN-generated fake images.The survey identifies DFFD as the only public database it knew of for this manipulation type.

B. Manipulation Detection

Attribute-manipulation detectors mainly use deep learning and achieve strong results, but performance depends on artifacts, frequency cues, and the evaluated generator or database.

  • Attribute-manipulation detection: 84.7% manipulation-detection accuracy was obtained by FakeSpoter using neuron-coverage features from deep face-recognition systems and FaceNet.The evaluation used real CelebA-HQ and FFHQ faces alongside InterFaceGAN- and StyleGAN-generated faces.
  • Attribute-manipulation detection: 99.4% accuracy was achieved by combining pixel co-occurrence matrices with a CNN on StarGAN-generated attribute manipulations.The fake dataset was generated from CelebA using StarGAN.
  • Attribute-manipulation detection: Frequency-domain detection reached 100% accuracy for StarGAN but fell to 50% for GauGAN.The authors associated the degradation with differences between GauGAN and the CycleGAN generator used in training.
  • Attribute-manipulation detection: A PRNU-based detector achieved an average 13.7% EER on a private database created with five mobile applications.The method fused spatial and spectral PRNU features extracted across image cells.
  • Attribute-manipulation detection: Deep learning forms the core of most attribute-manipulation detectors, with results generally close to 100% accuracy.The survey attributes these results mainly to GAN-fingerprint information in fake images.
  • Attribute-manipulation detection: Removing GAN fingerprints while retaining realistic appearance creates a challenge for advanced attribute-manipulation detectors.The survey identifies fingerprint removal as a broader difficulty for fake detection.

A. Manipulation Techniques and Public Databases

Expression swapping transfers facial expressions between subjects in videos or images, and detection methods combine spatial, temporal, and attention-based cues using limited public benchmarks.

  • Expression-swap techniques and databases: Face2Face and NeuralTextures replace one person’s facial expression with another’s in video, while FaceForensics++ is the only publicly available database identified for this area.FaceForensics++ extends FaceForensics and includes these video manipulations.
  • Expression-swap techniques and databases: Face2Face transfers 76 expression-related Blendshape coefficients from a source video to a target video while maintaining the target identity.NeuralTextures learns a target-person neural texture and modifies only the mouth expression in the described implementation.
  • Expression-swap techniques and databases: StarGAN and related GAN approaches can alter expression levels such as angry, happy, neutral, sad, surprised, and fearful in images.The survey also lists InterFaceGAN, UGAN, STGAN, and AttGAN as approaches improving quality or editing control.
  • Expression-swap detection: Detection performance was lower for NeuralTextures than Face2Face with mesoscopic and steganalysis features.The comparison was reported using FaceForensics++ fake videos.
  • Expression-swap detection: XceptionNet achieved results close to 100% on RAW-quality Face2Face and NeuralTextures videos, while accuracy degraded with lower video quality.The evaluation simulated processing conditions associated with social networks.
  • Expression-swap detection: 99.4% AUC and 3.4% EER were achieved by an attention-based detector on the expression-swap portion of DFFD.That portion of DFFD is based on FaceForensics++ data.
  • Expression-swap detection: Temporal and motion-based methods use recurrent convolutional networks or optical flow to exploit inter-frame discrepancies.The reported optical-flow system achieved 81.6% accuracy, while recurrent convolutional detection achieved 94.3% AUC for Face2Face.

VII. OTHER FACE MANIPULATION DIRECTIONS

Beyond the survey’s four main manipulation groups, the paper discusses face morphing, de-identification, and audio- or text-to-video synthesis as additional directions. These areas introduce distinct biometric, privacy, and detection challenges, with face morphing supported by emerging datasets and detectors.

  • Face Morphing: Face morphing creates artificial biometric samples resembling two or more individuals, potentially enabling verification against both identities.Unlike identity swaps, morphing primarily produces fake image samples rather than manipulated video.
  • Face Morphing: Face morphing differs from the survey’s four main manipulation types because it targets image-level biometric samples rather than video identity swaps.
  • Face Morphing: A morphing benchmark contains 1,800 real and morphed photographs from 150 subjects generated with six algorithms.
  • Face Morphing: Morphing detectors exploit blending-induced detail loss, sensor-pattern Fourier spectra, landmark differences, deep features, and face de-morphing.

B. Face De-Identification

Face de-identification removes or obscures facial identity to preserve privacy while potentially retaining non-identity attributes such as pose, expression, and illumination. Related audio- and text-to-video methods synthesize talking-face videos, but this area lacks public detection benchmarks.

  • Face De-Identification: Face de-identification removes identity information from face images or videos, ranging from blurring and pixelation to identity replacement that preserves other facial factors.
  • Face De-Identification: Generative deep neural networks, including GANs, have inspired newer de-identification methods for images and videos.
  • Face De-Identification: Some approaches disentangle protected information such as identity, gender, or expression before generating a modified face representation.
  • Audio-to-Video and Text-to-Video: Audio-to-video systems learn audio-to-mouth mappings or combine audio-visual features to synthesize realistic talking-face videos with lip synchronization and expression or pose variation.
  • Audio-to-Video and Text-to-Video: Text-to-video methods generate talking-face videos in which mouth movements are synchronized with newly specified words.
  • Audio-to-Video and Text-to-Video: No public databases or benchmarks are available for audio- and text-to-video fake detection, so studies typically synthesize in-house data.
  • Audio-to-Video and Text-to-Video: A detector exploits mismatches between mouth dynamics and spoken phonemes, especially failures to fully close the mouth for particular visemes.

VIII. CONCLUDING REMARKS

The survey consolidates face-manipulation techniques, datasets, detection methods, and benchmark results while identifying generalization, data scarcity, and robustness as central open issues. Detection is strong in controlled settings but degrades on newer or unseen manipulations, motivating broader benchmarks and alternative verification schemes.

  • Survey Scope: The survey organizes up-to-date manipulation types, techniques, public research databases, and representative detection benchmarks with their key results.
  • Generalization: Most detectors perform well under matched training and evaluation conditions, but social-media compression, resizing, noise, and evolving manipulations challenge generalization.
  • Future Directions: Feature- or score-level fusion is proposed as a way to adapt detectors across different scenarios by combining complementary information sources.
  • Future Directions: Social verification at capture time uses synchronized cameras to reach consensus and sign videos as true, offering an alternative to image- or video-only detection.
  • Face Synthesis: Face synthesis detectors often reach accuracies close to 100% because generated images retain specific GAN fingerprints, but removing those fingerprints remains challenging.
  • Identity Swap: Identity-swap comparisons remain difficult because methods use database-specific training, unseen-condition generalization is poor, and metrics and protocols differ.
  • Identity Swap: AUC results are close to 100% on first-generation databases but fall below 60% in most cases on Celeb-DF, a second-generation DeepFake database.
  • Attribute Manipulation: Attribute manipulation research is constrained by scarce public data, with only DFFD publicly available, and by a lack of standard experimental protocols.
Loading 2001.00179v3…