Backdoor Defense via Decoupling the Training Process

Huang, Kunzhe; Li, Yiming; Zhan, Qimin; Ren, Kui

doi:10.48550/arxiv.2202.03423

Cited by 9 publications

(15 citation statements)

References 25 publications

(46 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Poison Suppression. For poison suppression defense, most methods (e.g., DBD [22]) learn a backbone of a DNN model via self-supervised learning based on training samples without their labels to capture those suspicious training data during the training process. We tested the performance of DBD [22] defense against models attacked by IMC [33] with and without GRASP enhancement.…”

Section: Resilience To Backdoor Mitigationmentioning

confidence: 99%

“…For poison suppression defense, most methods (e.g., DBD [22]) learn a backbone of a DNN model via self-supervised learning based on training samples without their labels to capture those suspicious training data during the training process. We tested the performance of DBD [22] defense against models attacked by IMC [33] with and without GRASP enhancement. As shown in Table 4, the ASR of the DBD in the IMC* attacked models are higher than the IMC attacked models, indicating GRASP enhancement does not make the attack more easily to be mitigated by the DBD.…”

Section: Resilience To Backdoor Mitigationmentioning

confidence: 99%

“…In particular, our experiment shows that the GRASP enhancement does not decrease the effectiveness of the backdoor attacks such as DFST [16], AB [17], and DEFEAT [18] against the weight analysis-based detection. Finally, our study demonstrates that the effectiveness of GRASP against trigger inversion does not make the enhanced attacks more vulnerable to other backdoor mitigation or unlearning techniques, such as Finepurning [19], NAD [20], Gangsweep [21], DBD [22], and RAB [23]. Contributions.…”

Section: Introductionmentioning

confidence: 98%

See 2 more Smart Citations

Gradient Shaping: Enhancing Backdoor Attack Against Reverse Engineering

Zhu¹,

Tang²,

Tang³

et al. 2023

Preprint

View full text Add to dashboard Cite

Most existing methods to detect backdoored machine learning (ML) models take one of the two approaches: trigger inversion (aka. reverse engineer) and weight analysis (aka. model diagnosis). In particular, the gradientbased trigger inversion is considered to be among the most effective backdoor detection techniques, as evidenced by the TrojAI competition [1], Trojan Detection Challenge [2] and backdoorBench [3]. However, little has been done to understand why this technique works so well and, more importantly, whether it raises the bar to the backdoor attack. In this paper, we report the first attempt to answer this question by analyzing the change rate of the backdoored model around its trigger-carrying inputs. Our study shows that existing attacks tend to inject the backdoor characterized by a low change rate around trigger-carrying inputs, which are easy to capture by gradient-based trigger inversion. In the meantime, we found that the low change rate is not necessary for a backdoor attack to succeed: we design a new attack enhancement called Gradient Shaping (GRASP), which follows the opposite direction of adversarial training to reduce the change rate of a backdoored model with regard to the trigger, without undermining its backdoor effect. Also, we provide a theoretic analysis to explain the effectiveness of this new technique and the fundamental weakness of gradient-based trigger inversion. Finally, we perform both theoretical and experimental analysis, showing that the GRASP enhancement does not reduce the effectiveness of the stealthy attacks against the backdoor detection methods based on weight analysis, as well as other backdoor mitigation methods without using detection.

show abstract

Section: Resilience To Backdoor Mitigationmentioning

confidence: 99%

Section: Resilience To Backdoor Mitigationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 98%

See 1 more Smart Citation

Gradient Shaping: Enhancing Backdoor Attack Against Reverse Engineering

Zhu¹,

Tang²,

Tang³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Neural Cleanse [44], ANP [63], ABL [30], DBD [64] Activation Clustering [40], Spectral-signatures [41],…”

Section: Toolmentioning

confidence: 99%

Backdoor Pony: Evaluating backdoor attacks and defenses in different domains

Mercier¹,

Smolin²,

Sihlovec³

et al. 2023

SoftwareX

View full text Add to dashboard Cite

“…Backdoor Defenses. By exploiting certain properties of some backdoor attacks, various backdoor defenses [3,10,11,16,19,22,25,39,42,44,46,49,15] are also developed. As typical examples, Neural Cleanse [46] and SentiNet [7] propose to reverse engineer the backdoor triggers via searching for universal adversarial perturbations.…”

Section: Background and Related Workmentioning

confidence: 99%

Towards A Proactive ML Approach for Detecting Backdoor Poison Samples

Qi¹,

Xie²,

Mahloujifar³

et al. 2022

Preprint

View full text Add to dashboard Cite

Deep learning suffers from backdoor poisoning attacks. By injecting only a few poison samples into a training set, adversaries can easily embed stealthy backdoor into the trained models. In this work, we study poison samples detection for defending against backdoor poisoning attacks on deep neural networks (DNNs). A principled idea underlying prior arts on this problem is to utilize the backdoored models' distinguishable behaviors on poison and clean populations to distinguish between these two different populations themselves and remove the identified poison. Typically, many prior arts build their detectors upon a latent separability assumption, which states that backdoored models trained on the poisoned dataset will learn separable latent representations for backdoor and clean samples. Although such separation behaviors empirically exist for many existing attacks, there is no control on the separability and the extent of separation can vary a lot across different poison strategies, datasets, as well as the training configurations of backdoored models. Worse still, recent adaptive poison strategies can greatly reduce the "distinguishable behaviors" and consequently render most prior arts less effective (or completely fail). We point out that these limitations directly come from the passive reliance on some distinguishable behaviors that are not controlled by defenders. To mitigate such limitations, in this work, we propose the idea of active defense -rather than passively assuming backdoored models will have certain distinguishable behaviors on poison and clean samples, we propose to actively enforce the trained models to behave differently on these two different populations. Specifically, we introduce confusion training as a concrete instance of active defense. Confusion training separates poison and clean populations by introducing another poisoning attack to the already poisoned dataset, which actively decouples the benign correlations and leave backdoor correlations the only learnable patterns -consequently, only backdoor poison samples can be fitted, while clean samples are underfitted. In short, we literally invite a "defensive poison" in to fight the original backdoor poison we aim to cleanse. By extensive evaluations on both CIFAR10 and GTSRB, we show superiority of active defense across a diverse set of backdoor poisoning attacks.Preprint. Under review.

show abstract

Backdoor Defense via Decoupling the Training Process

Cited by 9 publications

References 25 publications

Gradient Shaping: Enhancing Backdoor Attack Against Reverse Engineering

Gradient Shaping: Enhancing Backdoor Attack Against Reverse Engineering

Backdoor Pony: Evaluating backdoor attacks and defenses in different domains

Towards A Proactive ML Approach for Detecting Backdoor Poison Samples

Contact Info

Product

Resources

About