Security of Facial Forensics Models Against Adversarial Attacks

Rong Huang, Fuming Fang, Huy H. Nguyen, Junichi Yamagishi, Isao Echizen

Abstract

Deep neural networks (DNNs) have been used in digital forensics to identify fake facial images. We investigated several DNN-based forgery forensics models (FFMs) to examine whether they are secure against adversarial attacks. We experimentally demonstrated the existence of individual adversarial perturbations (IAPs) and universal adversarial perturbations (UAPs) that can lead a well-performed FFM to misbehave. Based on iterative procedure, gradient information is used to generate two kinds of IAPs that can be used to fabricate classification and segmentation outputs. In contrast, UAPs are generated on the basis of over-firing. We designed a new objective function that encourages neurons to over-fire, which makes UAP generation feasible even without using training data. Experiments demonstrated the transferability of UAPs across unseen datasets and unseen FFMs. Moreover, we conducted subjective assessment for imperceptibility of the adversarial perturbations, revealing that the crafted UAPs are visually negligible. These findings provide a baseline for evaluating the adversarial security of FFMs.

Introduction

As facial manipulation technologies like DeepFakes become increasingly realistic, researchers have developed Deep Neural Network (DNN)-based Forgery Forensics Models (FFMs) to detect them. These models typically look for statistical inconsistencies, blending artifacts, or use autoencoders to classify and segment manipulated regions. However, the security of these forensic models themselves is rarely questioned. If a malicious actor can apply a tiny, invisible layer of noise to a DeepFake that tricks the FFM into classifying it as a "real" image, the entire forensic defense system collapses. This paper addresses the critical gap of evaluating whether digital forensics models can withstand targeted adversarial perturbations.

Methodology

The study mounts white-box and black-box attacks against a state-of-the-art multi-task autoencoder FFM, testing two distinct types of attacks:

Individual Adversarial Perturbations (IAPs)

IAPs are tailor-made noises for specific images. Using an iterative, gradient-based approach, the authors tweak the image to minimize an "activation loss." This allows them to not only flip the classification (fake to real) but also fabricate fake segmentation masks—for example, tricking the model into highlighting a randomly placed triangle as the "forged" region.

Overview of adversarial attacks on a multi-task FFM, showing classification flips and segmentation mask manipulation.

Universal Adversarial Perturbations (UAPs)

A UAP is a single, image-agnostic noise pattern that can fool the model on any image. Crucially, this is achieved without any training data. The authors designed a novel objective function that forces latent neurons in the target class's zone to "over-fire" (maximize their activation energy). By iteratively updating the noise to maximize this over-firing, they create a universal attack that contaminates the model's learned features.

Experimental Evaluation

The models and attacks were evaluated using the FaceForensics++ database, specifically the DeepFakes, Face2Face, and FaceSwap subsets.

IAP Effectiveness and Mask Hijacking

With just 20 iterations, IAPs dropped the target model's classification accuracy from ~90-94% down to 0% across all three datasets. Furthermore, the Intersection-over-Union (IoU) for the segmentation masks dropped to near zero. When actively trying to fabricate a fake segmentation mask, such as a random triangle, the IAPs achieved an IoU of ~48.5%, proving the model's localization can be controlled.

A fabricated segmentation mask (a triangle) generated by an IAP to mislead tampering localization.

UAP Performance and Transferability

The data-free UAPs reduced average classification accuracy from 91.98% to just 3.17%. These attacks also demonstrated significant transferability, successfully degrading the performance of unseen FFM architectures like MesoNet and ForensicTransfer by 19% to 24%. This suggests that the noise probes fundamental vulnerabilities in how these networks learn facial textures.

Human Perception and Imperceptibility

In a subjective test with 775 evaluators, UAPs scored 4.21/5 for visual quality (where the clean baseline was 4.49), proving the universal noise is virtually imperceptible to humans. IAPs scored slightly lower (3.45/5) due to localized distortions that the human eye can sometimes detect.

Conclusion

Forensic models are surprisingly fragile. Even advanced multi-task FFMs can be completely broken with minor, imperceptible perturbations. The effectiveness of data-free attacks and the ability to hijack segmentation results highlight an urgent need for the digital forensics community to integrate adversarial defense mechanisms to protect FFMs from being easily bypassed.

Summary updated

Abstract

Deep neural networks (DNNs) have been used in digital forensics to identify fake facial images. We investigated several DNN-based forgery forensics models (FFMs) to examine whether they are secure against adversarial attacks. We experimentally demonstrated the existence of individual adversarial perturbations (IAPs) and universal adversarial perturbations (UAPs) that can lead a well-performed FFM to misbehave. Based on iterative procedure, gradient information is used to generate two kinds of IAPs that can be used to fabricate classification and segmentation outputs. In contrast, UAPs are generated on the basis of over-firing. We designed a new objective function that encourages neurons to over-fire, which makes UAP generation feasible even without using training data. Experiments demonstrated the transferability of UAPs across unseen datasets and unseen FFMs. Moreover, we conducted subjective assessment for imperceptibility of the adversarial perturbations, revealing that the crafted UAPs are visually negligible. These findings provide a baseline for evaluating the adversarial security of FFMs.
BibTeX
@inproceedings{huang2020security,
  author = {Rong Huang and Fuming Fang and Huy H. Nguyen and Junichi Yamagishi and Isao Echizen},
  title = {{Security of Facial Forensics Models Against Adversarial Attacks}},
  booktitle = {ICIP},
  year = {2020},
  doi = {10.1109/icip40778.2020.9190678},
}
Figure