Towards Transferable Adversarial Attack Against Deep Face Recognition

Yaoyao Zhong, Weihong Deng

This paper, titled "Towards Transferable Adversarial Attack against Deep Face Recognition," presents a significant advancement in the field of AI security. It explores how to create "transferable" adversarial attacks—maliciously altered images that can fool a face recognition system even if the attacker has zero knowledge of that system’s internal architecture or training data.

TL;DR

The authors introduce DFANet (Dropout Face Attacking Network), a novel method that uses dropout in convolutional layers during attack generation to simulate a model ensemble. This approach significantly boosts the transferability of adversarial examples, allowing a single local model to generate attacks that successfully fool state-of-the-art face recognition systems and major commercial APIs (Amazon, Microsoft, Baidu, and Face++).

What Problem Does This Paper Address?

Deep Face Recognition (DFR) systems are now ubiquitous, used in everything from smartphone unlocking to border control. While highly accurate, they are vulnerable to adversarial attacks: small, human-imperceptible perturbations added to an image that cause a model to misidentify a person.

The authors focus on the most dangerous scenario: fully black-box transferable attacks.

  • White-box attacks assume the attacker knows everything about the target model (unrealistic).
  • Query-based black-box attacks require thousands of interactions with the target system, which are easily detected.
  • Transferable attacks are generated on a local "surrogate" model and "transferred" to the target.

The challenge is that an attack optimized for one model often fails on another because of differences in network architecture, training datasets, or loss functions. The authors aim to bridge this "transferability gap."

Methods

The paper makes two primary methodological contributions:

  1. Feature-Level vs. Label-Level Analysis: Most adversarial research focuses on object classification (e.g., "is this a cat?"). Face recognition is an "open-set" problem where the model produces a feature embedding (a mathematical representation of a face). The authors demonstrate that feature-level attacks (which aim to move one face's embedding closer to another's) are far more transferable than traditional label-level attacks.
  2. DFANet (Dropout Face Attacking Network): To prevent an attack from "overfitting" to the specific quirks of the local surrogate model, the authors propose using dropout in the convolutional layers during the attack generation phase.
  3. By randomly "dropping" neurons at each iteration of the attack, they effectively force the perturbation to be effective against a different "sub-network" every time.
  4. This creates an ensemble-like effect using only a single surrogate model, making the resulting adversarial image much more robust and likely to transfer to unknown target models.

Datasets

The study utilizes several mainstream DFR datasets for training surrogate models and evaluation:

  • Training/Surrogates: CAISA-WebFace, MS-Celeb-1M, VGGFace2, and IMDb-Face.
  • Evaluation: LFW (Labeled Faces in the Wild), MEDS, and PaSC.
  • New Benchmark: The authors introduce TALFW (Transferable Adversarial Labeled Faces in the Wild). This is a modified version of the famous LFW dataset where images have been subtly altered using DFANet. It serves as a new "stress test" for the robustness of face recognition systems.

Key Insights and Ideas

  • The "Ensemble" Shortcut: Training a real ensemble of multiple deep models is computationally expensive. DFANet provides the benefits of an ensemble (diversity and generalization) at the cost of a single model by leveraging the randomness of dropout.
  • Feature Space is Universal: While different models might have different classification layers, the underlying "feature space" they learn for faces is remarkably similar. Targeting the embedding distance is the most effective way to exploit this shared logic.
  • Overfitting is the Enemy: Standard adversarial attacks are too "fine-tuned" to the surrogate. Introducing architectural noise (via dropout) and input noise (via image transformations) is essential for a successful black-box attack.

Key Results

  • Superior Hit Rates: When combined with momentum and diverse input methods, DFANet achieved nearly 90% success rates in transferring attacks between different architectures (e.g., from a ResNet-50 surrogate to a SENet-50 target).
  • Fooling the Giants: The authors successfully attacked four major commercial Face Recognition APIs: Amazon, Microsoft, Baidu, and Face++. In one example, a pair of images of different people—initially rated at 17.8% similarity—was judged to be the same person with 78.3% similarity after the DFANet perturbation was added.
  • TALFW Impact: On the new TALFW benchmark, the accuracy of the state-of-the-art ArcFace model plummeted from 99.82% to 63.45%, highlighting a massive vulnerability in current "secure" systems.

Limitations and Future Work

  • Defensive Trade-offs: The authors tested several defenses (JPEG compression, Gaussian blur, adversarial training). While these methods improve robustness, they often slightly degrade the model's accuracy on clean, non-adversarial images.
  • The "Cat and Mouse" Game: While DFANet is powerful, the authors acknowledge that as defense techniques evolve (like more sophisticated adversarial training), attack methods will need to become even more diverse.
  • TALFW Availability: The authors have released the TALFW database to the research community to encourage the development of more robust DFR models that can withstand these transferable threats.

Summary updated

Abstract

Face recognition has achieved great success in the last five years due to the development of deep learning methods. However, deep convolutional neural networks (DCNNs) have been found to be vulnerable to adversarial examples. In particular, the existence of transferable adversarial examples can severely hinder the robustness of DCNNs since this type of attacks can be applied in a fully black-box manner without queries on the target system. In this work, we first investigate the characteristics of transferable adversarial attacks in face recognition by showing the superiority of feature-level methods over label-level methods. Then, to further improve transferability of feature-level adversarial examples, we propose DFANet, a dropout-based method used in convolutional layers, which can increase the diversity of surrogate models and obtain ensemble-like effects. Extensive experiments on state-of-the-art face models with various training databases, loss functions and network architectures show that the proposed method can significantly enhance the transferability of existing attack methods. Finally, by applying DFANet to the LFW database, we generate a new set of adversarial face pairs that can successfully attack four commercial APIs without any queries. This TALFW database is available to facilitate research on the robustness and defense of deep face recognition.
BibTeX
@article{zhong2021towards,
  author = {Yaoyao Zhong and Weihong Deng},
  title = {{Towards Transferable Adversarial Attack Against Deep Face Recognition}},
  journal = {IEEE Trans. Inf. Forensics Secur.},
  year = {2021},
  url = {https://openreview.net/forum?id=7h_9cthbslI},
}

Loading related content...

Figure