Temporal Consistency Constrained Transferable Adversarial Attacks with Background Mixup for Action Recognition

Ping Li, Jianan Ni, Bo PangIJCAI 2025 · DOI

TL;DR

This paper introduces a novel adversarial attack method called BMTC for action recognition models. BMTC enhances the transferability of adversarial attacks by using a "background mixup" technique to create more robust adversarial examples and a temporal consistency loss to ensure stable attack directions across video frames. This approach significantly boosts attack success rates against black-box models, even those with different architectures.

What Problem Does This Paper Address?

Action recognition models, crucial for applications like video surveillance and robotics, are increasingly relying on deep neural networks. However, these models are vulnerable to adversarial examples – subtly modified inputs that fool the model into making incorrect predictions. A significant challenge in crafting these attacks is their transferability: an adversarial example generated for one model (the surrogate) might not effectively fool a different model (the target), especially if their decision boundaries differ. Existing methods often assume similar decision boundaries, limiting their effectiveness in real-world scenarios where target model details are unknown. Furthermore, the uncertainty in attack direction can lead to gradient oscillations, weakening the attack. This paper aims to overcome these limitations by developing a transferable adversarial attack that is less dependent on surrogate-target model similarity and ensures a consistent, strong attack direction.

Methods

The proposed Background Mixup-induced Temporal Consistency (BMTC) attack method comprises two main components:

  1. Background Adversarial Mixup (BAM) Module: This module addresses the reliance on similar decision boundaries. Instead of directly perturbing the input, it uses a model-agnostic input transformation. Specifically, it randomly samples background frames from different action categories and mixes them with the clean video frames. To select the most effective background frame for mixing, it employs reinforcement learning, optimizing a reward function that considers:
  2. Attack Reward: How well the admixed sample fools the surrogate model.
  3. Transfer Reward: The attack's success against multiple target models.
  4. Temporal Background Consistency Reward: Encourages the use of similar background frames for adjacent video frames to maintain temporal coherence.
  5. Background-induced Temporal Gradient Enhancement (BTG) Module: This component tackles the problem of uncertain attack directions. It introduces two adversarial losses:
  6. Background Attack Loss: Guides the surrogate model to generate adversarial examples that fool the model into predicting the selected background category, leveraging the semantic correlation between actions and their backgrounds.
  7. Temporal Gradient Consistency Loss: Encourages the gradients of adversarial perturbations to be consistent across consecutive frames. This ensures a stable and progressive attack direction, strengthening the overall attack.

These components are integrated into a framework that first generates an admixed sample using BAM and then refines it using PGD with the BTG losses to produce the final adversarial example.

Datasets

The paper evaluates its method on the following datasets:

  • UCF101: A widely used dataset for action recognition.
  • Kinetics-400: Another large-scale dataset for action recognition.
  • ImageNet: A benchmark dataset for image classification, used to demonstrate cross-modal transferability.

Key Insights and Ideas

  • Leveraging Background Semantics: The core idea that actions are often associated with specific backgrounds (e.g., surfing in the sea) is ingeniously used to create adversarial examples. By mixing in backgrounds from other categories, the model can be misled.
  • Reinforcement Learning for Optimal Mixup: Using RL to select the "best" background frame for mixing is a clever way to automate and optimize this process, going beyond simple random mixing.
  • Temporal Consistency is Key for Videos: The paper highlights that simply attacking frames independently isn't optimal for videos. Ensuring consistent attack directions across frames leads to a more potent and stable attack.
  • Bridging the Surrogate-Target Gap: The BAM module's ability to reduce reliance on similar decision boundaries is crucial for practical black-box attacks.

Key Results

The BMTC attack method demonstrates significant improvements in Attack Success Rate (ASR) across various action recognition models (CNNs like NL, SlowFast, TPN, and ViTs like VTN, TimeSformer) and datasets (UCF101, Kinetics-400).

  • UCF101: BMTC achieved ASRs significantly higher than baseline methods like PGD and TT, often by substantial margins (e.g., gains of 9.0% and 21.0% over TT in specific scenarios).
  • Kinetics-400: Similar improvements were observed, showcasing the method's effectiveness on larger datasets.
  • Cross-Modal Transfer: The method also showed strong performance in transferring attacks from image models to video models.
  • Efficiency: BMTC achieves higher ASR with significantly lower computational cost (GFLOPs) compared to methods like Temporal Translation (TT), offering a better trade-off between performance and efficiency.
  • Qualitative Results: Visualizations show that the perturbations introduced by BMTC are almost imperceptible to humans, making them more stealthy.

Limitations and Future Work

  • Hyperparameter Sensitivity: The paper notes that the performance is sensitive to the choice of hyperparameters like $\gamma$ (mixup ratio) and $\beta$ (temporal gradient consistency loss weight), requiring careful tuning.
  • Human Imperceptibility: While the paper aims for imperceptible noise, the ablation study on $\epsilon$ (perturbation magnitude) suggests that larger values of $\epsilon$ boost performance but might violate the human-imperceptibility constraint.
  • Future Work: The paper implicitly suggests that further exploration into optimizing the RL reward function and exploring different background selection strategies could yield even better results. The success on both video and image domains also opens avenues for more sophisticated cross-modal attacks.

Summary updated

BibTeX
@inproceedings{li2025temporal,
  author = {Ping Li and Jianan Ni and Bo Pang},
  title = {{Temporal Consistency Constrained Transferable Adversarial Attacks with Background Mixup for Action Recognition}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/158},
}
Figure