This work comprehensively investigates computation-efficient strategies to speed up latent adversarial training from two complementary perspectives, and reduces per-step adversarial-training FLOPs by 48.1% while requiring only 0.0118% trainable parameters.
Abstract
Adversarial training is one of the most effective defenses against adversarial attacks, yet the computational cost remains prohibitive at modern scales, especially for large language models (LLMs). While existing mitigation strategies, e.g., latent adversarial training (LAT), have been developed, they still incur a high computational cost. In this work, we comprehensively investigate computation-efficient strategies to speed up LAT from two complementary perspectives: (1) Defense-side optimization: We explore the representation fine-tuning (ReFT) within LAT, and reveal a potential issue if there is a mismatch on which tokens to apply ReFT and the attack. (2) Attack-side optimization: When computing adversarial attacks in each LAT iteration, we extract only the relevant circuits from the LLM to construct a lightweight surrogate model, avoiding the computation in the forward-backward passes through the full model during the attack generation. For both perspectives, we provide theoretical justifications and numerical evidence to illustrate the effectiveness of the proposed strategies. Ultimately, compared to standard LAT with full fine-tuning, our method on average reduces per-step adversarial-training FLOPs by 48.1% while requiring only 0.0118% trainable parameters.
This work proposes BMAT (Bilevel-Minimax Adversarial Transfer), an integrated bottom-up solver that combines a Soft Weight Modulator and an Implicit Gradient Approximator to enable ternary coupling among initialization, surrogate adaptation, and perturbation optimization.
Adversarial training (AT) is a widely adopted defense against adversarial attacks, but its multi-step optimization process for generating adversarial examples leads to substantial computational overhead. To mitigate this, various single-step adversarial training methods have been proposed. However, these models often suffer from catastrophic overfitting under larger perturbations and exhibit degraded robustness. The core issue is that certain single-step adversarial examples, although successfully learned and correctly classified, fail to expose the true vulnerabilities of models. We refer to these misleading examples as “fakers”. Specifically, we find that fakers exhibit three distinct characteristics compared to standard adversarial examples: 1) they unexpectedly degrade the model robustness rather than improve it; 2) they make it harder for the model to learn their robust features; and 3) they show significantly greater divergence from their clean counterparts. These observations motivate us to proactively reduce the impact of fakers during training. To this end, we propose the Faker-Alleviating Single-step adversarial Training method (FAST), a general and effective framework designed to enhance both accuracy and robustness. Concretely, FAST consists of two main components. First, it dynamically adjusts the label-smoothing level for adversarial examples according to their learning difficulty, making fakers easier for the model to learn. Second, it introduces an auxiliary sample with a weak adversarial effect, derived from the single-step adversarial example, which is used to dynamically ease the alignment with clean data and stabilize the optimization process. We demonstrate the effectiveness of FAST through extensive experiments, showing that our method achieves superior clean accuracy and robustness against various types of adversarial attacks. The code is available at https://github.com/mesunhlf/FAST.
Lifeng Huang, Yuquan Lin, Chen Wan et al.· IEEE Transactions on Informa...· 0 citations
Vision Transformers (ViTs) remain vulnerable to localized adversarial attacks, e.g., adversarial patches, while recent test-time defenses mitigate them by suppressing image tokens with abnormally high attention scores. These defenses exploit a strong coupling between attention and adversarial effectiveness: adversarial tokens often need to attract substantial attention to influence the prediction. We introduce adversarial decoys, independently optimized image patches that redirect the attention, and therefore related defenses, toward selected target tokens. Rather than jointly optimizing misclassifications and defense evasion, our approach decouples the two objectives: the original adversarial region induces the incorrect prediction, while a separate decoy manipulates the attention ranking used by the defense. A layer-wise objective increases target-token attention and promotes these tokens above competing non-target ones. Since the decoy is optimized independently of the underlying attack, the method is attack-agnostic and can be easily integrated with any existing adversarial patch attack. Experiments on ImageNet across multiple ViT architectures and attacks show that decoys can redirect high attention scores away from the true adversarial region while preserving much of the attack effectiveness. These results reveal a fundamental limitation of using attention magnitude as an indicator of adversarial relevance.
Giulia Marchiori Pietrosanti, Giulio Rossolini, Giorgio C. Buttazzo· 0 citations
Vision Transformers (ViTs) increasingly rely on input-adaptive inference, such as token pruning and early halting, to meet energy and latency budgets. This survey examines a recent class of adversarial efficiency degradation attacks that target these mechanisms to increase computation without necessarily degrading accuracy. We unify and compare two representative attacks, SlowFormer (a universal adversarial patch) and DeSparsify (per-image perturbations), across three popular token-pruning frameworks: A-ViT, ATS, and AdaViT. We standardize reporting using GFLOPs, accuracy loss, and an Attack Success (AS) metric that measures how much of the model's compute savings the attack takes away. Understanding these attacks is crucial for designing countermeasures that not only mitigate risk but also remain lightweight, since deployment often occurs in low-power settings such as mobile or embedded devices. To organize our analysis, we focus on three questions: how input-adaptive optimizations (e.g., token pruning and early halting) create attack surfaces for efficiency degradation; how such attacks operate in practice and which optimizations are most vulnerable; and which defenses exist today and whether they meaningfully restore efficiency under attack.
Anadi Goyal, Nandish Chattopadhyay, Anupam Chattopadhyay et al.· 0 citations
Machine learning models are increasingly adapted in various domains. However, adversarial examples pose a significant threat to the reliable deployment of these models. In recent years, some powerful adversarial example attacks have been proposed for the fast and query-efficient generation of adversarial examples, even in black-box scenarios, highlighting the need for scalable, low-cost, and powerful defenses. In this work, we present two contributions to the domain of black-box adversarial example attacks and defenses. First, we propose Random Logit Scaling (RLS), a randomization-based defense against black-box score-based adversarial example attacks. RLS is a plug-and-play, post-processing defense that can be implemented on top of any existing ML model with minimal effort. The idea behind RLS is to confuse an attacker by outputting falsified scores resulting from randomly scaled logits while maintaining the model accuracy. We show that RLS significantly reduces the success rate of state-of-the-art black-box score-based attacks while preserving the accuracy and minimizing confidence score distortion compared to state-of-the-art randomization-based defenses. Second, we introduce a novel adaptive attack against AAA, a SOTA non-randomized black-box defense against black-box score-based attacks that also modifies output logits to confuse attackers, demonstrating its vulnerability against adaptive attacks.
Hamid Dashtbani, Mehdi Dousti Gandomani, A. M. Sadeghzadeh· 0 citations
Adversarial training has emerged as a leading strategy for enhancing the robustness of machine learning models against adversarial attacks. Its effectiveness often wanes when faced with unseen adversarial examples, resulting in suboptimal robust generalization. To address this issue, we introduce a novel energy-based optimization strategy to improve the robust generalization by incorporating the principles of energy-based models. Our framework models the energy of natural and adversarial examples, where natural samples are assigned to lower energy and adversarial samples to higher energy. During the inference phase, the influence of adversarial perturbation can be alleviated by energy minimization. Theoretically, we show that the proposed energy-based optimization strategy yields a tighter robust-generalization bound through an explicit energy-discrepancy term; this analysis provides an explanatory bound and should not be interpreted as certified robustness. Empirically, a series of evaluations provide evidence for the efficacy of the proposed methodology under the specified threat models and evaluation protocols, showing strong and competitive performance across three extensively utilized datasets. Specifically, EM-AT achieves 77.71% standard-AA robustness on CIFAR-10 and remains highly competitive under comparable lightweight settings. The source codes are available at https://github.com/LitterQ/EM-AT.