TL;DR
This paper introduces a novel adversarial attack method called Beneficial Perturbation Feature Augmentation (BPFA) that significantly improves the transferability of adversarial face recognition attacks. By creating "hard models" that increase the difficulty for a face recognition system, BPFA helps adversarial examples bypass more models, thus exposing critical security vulnerabilities.
What Problem Does This Paper Address?
Face recognition (FR) systems are increasingly deployed in critical infrastructure, from airport security to financial transactions. However, these systems are vulnerable to adversarial examples – subtly modified images that fool the model into misidentifying individuals. A major concern is the transferability of these attacks: adversarial examples crafted against one FR model can often fool other, unseen models. This poses a significant threat to national security and digital governance, as attackers can exploit these vulnerabilities without prior knowledge of the target system. While data augmentation with "hard samples" is known to improve model generalization, its application to enhancing adversarial attack transferability remained unexplored. This paper bridges this gap by proposing a method to create "hard models" to augment the adversarial attack process.
Methods
The core of the proposed method is the Beneficial Perturbation Feature Augmentation Attack (BPFA). BPFA leverages the concept of "hard models," which are defined as models that produce higher losses for the same input compared to normal models. This concept is inspired by "hard samples" in training, which are known to improve model generalization.
BPFA works by:
- Generating Hard Models: During the adversarial attack process, BPFA introduces "beneficial perturbations" to specific feature maps within a surrogate FR model. These perturbations are designed to increase the model's loss, effectively creating a "harder" version of the model.
- Augmenting the Surrogate Model: By repeatedly generating these hard models, BPFA effectively augments the "training set" for crafting adversarial examples, preventing them from overfitting to a single surrogate model.
- Gradient Utilization: To reduce computational overhead, BPFA uses gradients recorded from a previous backpropagation step to simulate the gradients needed for crafting beneficial perturbations.
- Minimax Optimization: The attack operates under a minimax framework where adversarial perturbations on the input image aim to decrease the loss, while beneficial perturbations on feature maps aim to increase it.
Datasets
The experiments were conducted on two standard face recognition datasets:
- LFW (Labeled Faces in the Wild): A widely used dataset for unconstrained face recognition.
- CelebA-HQ: A dataset containing high-quality face images.
For evaluation, a subset of 1000 face pairs from each dataset was used, consisting of negative pairs (images of different individuals).
Key Insights and Ideas
- "Hard Models" for Attack Transferability: The paper introduces the novel concept of "hard models" as a way to improve adversarial attack transferability, drawing an analogy to "hard samples" in model training.
- Beneficial Perturbations as a Countermeasure: The use of "beneficial perturbations" to increase model loss is a counter-intuitive but effective strategy for creating harder models that enhance adversarial attack success.
- Semantic Information in Perturbations: Visualizations suggest that beneficial perturbations added to feature maps can contain semantic information, potentially related to contours or important facial features, which contributes to their effectiveness.
- BPFA Enhances Existing Attacks: BPFA doesn't replace existing attack methods but rather augments them, significantly boosting their black-box transferability without harming white-box performance.
Key Results
- BPFA significantly improves the attack success rate (ASR) of various baseline adversarial attacks on both LFW and CelebA-HQ datasets. For instance, when combined with FMDN (a state-of-the-art attack), BPFA consistently boosted ASR across different victim models (FaceNet, MF, IRSE50, IR152).
- The method demonstrated effectiveness even against adversarial robust models, showing that BPFA can still enhance attacks on models trained to resist them.
- Experiments on dodging attacks also showed improved performance when BPFA was applied.
- Ablation studies indicated that the effectiveness of BPFA is sensitive to the step size of beneficial perturbations (η), with an optimal range for maximizing attack performance.
- The storage overhead introduced by BPFA is minimal, making it a practical addition to existing attack methods.
Limitations and Future Work
- Computational Overhead: While BPFA uses simulated gradients to save computation, it still requires recording gradients from previous steps, which adds some storage overhead. The paper notes this is acceptable in scenarios where storage is not a primary constraint.
- Optimal Hyperparameter Tuning: The effectiveness of BPFA is influenced by hyperparameters like the step size (η) for beneficial perturbations. Finding the optimal values for different scenarios might require further exploration.
- Interpretability of Semantic Information: While semantic information was observed in beneficial perturbations, a deeper understanding of its exact role and how it contributes to transferability could be a future research direction.
- Broader Applicability: The paper focuses on face recognition. Future work could explore the applicability of BPFA to other domains of deep learning where adversarial robustness is a concern.
Summary updated