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.
Introduction
Deep Face Recognition (DFR) systems have become a cornerstone of modern security, powering applications ranging from mobile device authentication to border control. Despite their high accuracy, these systems are susceptible to adversarial attacks—subtle, human-imperceptible modifications to images that trigger incorrect classifications. The most critical threat comes from transferable adversarial attacks, which are generated on a local surrogate model and successfully fool a different, unknown target system. This "black-box" capability is particularly dangerous as it requires no internal knowledge of the target architecture or training data.
The Transferability Gap in Face Recognition
Standard adversarial attacks often "overfit" to the specific parameters and quirks of the local surrogate model used to generate them. When these attacks are applied to a target model with a different architecture, loss function, or training set, they often fail.
In the context of DFR, which is an "open-set" problem, models produce feature embeddings rather than simple labels. The authors demonstrate that feature-level attacks—which aim to manipulate the mathematical distance between embeddings in feature space—are significantly more transferable than traditional label-level attacks used in object classification. This suggests that while classification layers vary, the underlying feature space learned by different DFR models is remarkably similar.
DFANet: Simulating Ensembles via Dropout
To bridge the transferability gap, the researchers propose the Dropout Face Attacking Network (DFANet). The core innovation lies in applying dropout within the convolutional layers during the attack generation phase.
By randomly dropping neurons at each iteration of the attack, DFANet forces the perturbation to be effective against a slightly different "sub-network" every time. This process creates an ensemble-like effect using only a single surrogate model. This architectural noise prevents the attack from becoming too fine-tuned to the surrogate, resulting in a more robust and transferable adversarial image.
Experimental Results and Commercial Impact
The effectiveness of DFANet was validated through extensive testing across various architectures (e.g., ResNet-50, SENet-50) and training datasets. Key findings include:
- High Transferability: When combined with momentum and diverse input methods, DFANet achieved success rates of nearly 90% when transferring attacks between different network architectures.
- Fooling Commercial APIs: The method successfully compromised four major commercial face recognition services: Amazon, Microsoft, Baidu, and Face++. In one instance, a pair of images of different individuals, initially rated at 17.8% similarity, was judged to be the same person with 78.3% similarity after the DFANet perturbation.
- The TALFW Benchmark: The authors introduced the Transferable Adversarial Labeled Faces in the Wild (TALFW) database. On this new benchmark, the accuracy of the state-of-the-art ArcFace model dropped from 99.82% to 63.45%, highlighting a significant vulnerability in current production-grade systems.
Conclusion
DFANet represents a significant step forward in understanding and exploiting the vulnerabilities of deep face recognition. By leveraging dropout to simulate model ensembles, the researchers have demonstrated that even sophisticated, commercial-grade systems are vulnerable to black-box attacks. The release of the TALFW database provides the research community with a vital tool for developing more robust defense mechanisms and improving the overall security of DFR technology.
Summary updated