Understanding Impacts of Task Similarity on Backdoor Attack and Detection

Di Tang, Rui Zhu, XiaoFeng Wang, Haixu Tang, Yi ChenarXiv 2022 · DOI · arXiv

Abstract

This paper introduces Task Similarity Analysis (TSA), a formal framework demonstrating that a backdoor attack's stealthiness is intrinsically linked to the similarity between the "backdoor task" and the model's "primary task." By quantifying this relationship using a novel distance metric, the authors develop the TSA attack. This optimization-based approach generates ultra-stealthy backdoors that maintain high attack success rates (ASR) even at extremely low poisoning rates, effectively evading state-of-the-art detection methods.

Introduction

Backdoor attacks pose a significant security threat to machine learning models, where an adversary embeds a hidden trigger to induce specific misbehaviors. While many attacks focus on visual subtlety, automated detectors often analyze model weights or internal activations, rendering visual stealth insufficient. This research addresses the lack of a fundamental mathematical understanding of backdoor stealth by asking whether the "distance" between a benign task and a backdoor task determines detectability.

Theoretical Framework and Methodology

The authors treat a backdoored model as performing two concurrent tasks: a primary task (e.g., recognizing traffic signs) and a backdoor task (e.g., misclassifying signs with a specific trigger). They define a distance metric, $d_{\mathcal{H}-W1}$, combining Wasserstein-1 distance and $\mathcal{H}$-divergence to quantify the difference between these tasks. A key innovation is the $\alpha$-backdoor metric, a normalized measure of task similarity that allows for comparison across different poisoning rates.

The resulting TSA attack optimizes the trigger to map inputs near the benign model's classification boundary while ensuring the trigger is "easy to learn." It regularizes training to minimize output drift between backdoored and benign models, effectively hiding the backdoor within the primary task's logic.

Visual comparison of triggers and task similarity metrics across different backdoor attacks.

Experimental Evaluation

The TSA attack was evaluated on CIFAR10, GTSRB, ImageNet, and VGGFace2 against five prior attacks: BadNet, SIG, WB (Wasserstein Backdoor), CB (Composite Backdoor), and IAB (Input-Aware Backdoor). A core finding is the strong positive correlation (0.9777) between mathematical task distance and actual detectability. If the tasks are mathematically similar, the attack becomes fundamentally harder for detectors to distinguish from benign behavior.

Performance comparison showing Attack Success Rate (ASR) and task similarity metrics across various poisoning rates.

Key Results

The TSA attack demonstrates superior evasion capabilities, reducing detection accuracy to near-random levels (50-60%) for six state-of-the-art detectors, including K-ARM and MNTD. Notably, it maintains an ASR of 79.07% at a poisoning rate as low as 0.005%, significantly outperforming traditional attacks like BadNet, which drops to 38.85% at the same rate. The results validate that true stealth requires balancing input similarity (how likely triggers appear in the original distribution) and output similarity (how similar the internal logic is for both tasks).

Conclusion

TSA provides a rigorous theoretical foundation for understanding backdoor stealth. By minimizing the "drift" in both inputs and outputs simultaneously, the TSA attack exposes blind spots in current detection methods. While highly effective, future work is needed to extend the framework to other domains like Natural Language Processing (NLP). The study suggests that future defenses must evolve to simultaneously monitor both input and output anomalies to counter similarity-optimized attacks.

Summary updated

BibTeX
@misc{tang2022understanding,
  author = {Di Tang and Rui Zhu and XiaoFeng Wang and Haixu Tang and Yi Chen},
  title = {{Understanding Impacts of Task Similarity on Backdoor Attack and Detection}},
  year = {2022},
  eprint = {2210.06509},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2210.06509},
}
Figure