TL;DR
This paper introduces MUAP, a novel adversarial attack method that can fool person re-identification (Re-ID) systems even when the attacker has limited knowledge of the target model. MUAP achieves this by disrupting the entire similarity ranking list and by being "model-insensitive," meaning it works across different Re-ID models and datasets, making current Re-ID systems surprisingly vulnerable.
What Problem Does This Paper Address?
Person re-identification (Re-ID) systems are crucial for tasks like surveillance and security, aiming to identify individuals across different cameras. While deep learning has significantly improved Re-ID performance, these systems are vulnerable to adversarial attacks. Existing attacks often require specific knowledge of the target model or generate perturbations for each individual image, limiting their real-world applicability. Furthermore, these attacks typically focus on classification tasks, which have clear decision boundaries, unlike Re-ID, which relies on similarity rankings. This paper addresses the challenge of creating a "universal" adversarial perturbation (UAP) that can effectively attack Re-ID systems in a more realistic, black-box setting, where the attacker has minimal information about the target model and its architecture.
Methods
The authors propose a "More Universal Adversarial Perturbation" (MUAP) method with two key components:
- List-wise Attack Objective (LAP): Instead of trying to misclassify an image (as in classification attacks), MUAP aims to disrupt the entire similarity ranking list. This is achieved by optimizing an "Average Precision" (AP) loss function, which is adapted to be differentiable for end-to-end training. This objective directly targets the core of Re-ID by making it harder for the system to correctly rank similar individuals.
- Model-Insensitive (MI) Regularization: To ensure the attack works across different Re-ID models (i.e., is "model-insensitive"), a regularization term is introduced. This term encourages the generated perturbations to have a natural image gradient distribution, preventing the attack from overfitting to a specific model's architecture. This makes the perturbation more transferable to unseen models and datasets.
The overall optimization process involves iteratively updating the perturbation vector using stochastic gradient descent, with constraints to keep the perturbation magnitude small.
Datasets
The paper evaluates MUAP on several widely used Re-ID datasets:
- Market-1501
- DukeMTMC-reID
- MSMT17
- MARS (for video-based Re-ID)
These datasets are standard benchmarks for person re-identification research.
Key Insights and Ideas
- Re-ID is fundamentally different from classification for adversarial attacks: The lack of clear decision boundaries in Re-ID means attacks need to disrupt ranking, not just labels.
- Universal perturbations can be highly effective for Re-ID: The concept of a single perturbation that works across many images is powerful, especially when combined with Re-ID specific objectives.
- Model-insensitivity is crucial for real-world attacks: Attacks that only work on the exact model they were trained on are less threatening. MUAP's MI regularization is key to overcoming this limitation.
- Naturalness of perturbations matters for transferability: The MI regularization helps create perturbations that look more natural, which paradoxically makes them more effective at fooling different models.
- Current Re-ID systems are surprisingly fragile: The results demonstrate that even state-of-the-art Re-ID models can be significantly degraded by MUAP, even in challenging cross-model and cross-dataset scenarios.
Key Results
- MUAP significantly outperforms existing state-of-the-art attack methods in black-box scenarios (cross-model and cross-dataset attacks), achieving mean Average Precision (mAP) drop rates (mDR) of over 90% in many cases.
- In white-box scenarios (same model, same dataset), MUAP also achieves competitive or superior performance compared to other methods.
- The MI regularization component is shown to dramatically improve the transferability of UAPs across different CNN architectures and datasets, increasing mDR from around 30% to over 75% in some cross-model scenarios.
- The proposed list-wise attack objective (LAP) consistently outperforms traditional classification-based attack methods for Re-ID.
- The attack is effective against various Re-ID frameworks, including part-based, attention-based, and multi-clue models, significantly degrading their accuracy.
Limitations and Future Work
- Computational Cost: While UAP is generally efficient, training the MUAP can still be computationally intensive.
- Defense Mechanisms: The paper highlights the vulnerability of current Re-ID systems and implicitly suggests the need for more robust model designs. Future work could focus on developing effective defenses against such universal adversarial attacks.
- Exploration of MI Regularization Forms: The paper explores different forms of MI regularization (l1, l2, and combinations) and finds that the l2 form or a combination of l1 and l2 is generally superior for cross-model attacks. Further fine-tuning of these parameters could potentially yield even better results.
- Real-world Deployment: While the paper demonstrates effectiveness in simulated environments, further research could explore the practical challenges of deploying such attacks in real-world scenarios, considering factors like image compression and varying lighting conditions.
Summary updated