Moderating the Generalization of Score-based Generative Model

Score-based Generative Models (SGMs) have demonstrated remarkable generalization abilities, e.g. generating unseen, but natural data. However, the greater the generalization power, the more likely the unintended generalization, and the more dangerous the abuse. Research on moderated generalization in SGMs remains limited. To fill this gap, we first examine the current 'gold standard' in Machine Unlearning (MU), i.e., re-training the model after removing the undesirable training data, and find it does not work in SGMs. Further analysis of score functions reveals that the MU 'gold standard' does not alter the original score function, which explains its ineffectiveness. Based on this insight, we propose the first Moderated Score-based Generative Model (MSGM), which introduces a novel score adjustment strategy that redirects the score function away from undesirable data during the continuous-time stochastic differential equation process. Extensive experimental results demonstrate that MSGM significantly reduces the likelihood of generating undesirable content while preserving high visual quality for normal image generation. Albeit designed for SGMs, MSGM is a general and flexible MU framework that is compatible with diverse diffusion architectures (SGM and DDPM) and training strategies (re-training and fine-tuning), and enables zero-shot transfer of the pre-trained models to downstream tasks, e.g. image inpainting and reconstruction. The code will be shared upon acceptance.

TASAR: Transferable Attack on Skeletal Action Recognition

Skeletal sequences, as well-structured representations of human behaviors, are crucial in Human Activity Recognition (HAR). The transferability of adversarial skeletal sequences enables attacks in real-world HAR scenarios, such as autonomous driving, intelligent surveillance, and human-computer interactions. However, existing Skeleton-based HAR (S-HAR) attacks exhibit weak adversarial transferability and, therefore, cannot be considered true transfer-based S-HAR attacks. More importantly, the reason for this failure remains unclear. In this paper, we study this phenomenon through the lens of loss surface, and find that its sharpness contributes to the poor transferability in S-HAR. Inspired by this observation, we assume and empirically validate that smoothening the rugged loss landscape could potentially improve adversarial transferability in S-HAR. To this end, we propose the first Transfer-based Attack on Skeletal Action Recognition, TASAR. TASAR explores the smoothed model posterior without re-training the pre-trained surrogates, which is achieved by a new post-train Dual Bayesian optimization strategy. Furthermore, unlike previous transfer-based attacks that treat each frame independently and overlook temporal coherence within sequences, TASAR incorporates motion dynamics into the Bayesian attack gradient, effectively disrupting the spatial-temporal coherence of S-HARs. To exhaustively evaluate the effectiveness of existing methods and our method, we build the first large-scale robust S-HAR benchmark, comprising 7 S-HAR models, 10 attack methods, 3 S-HAR datasets and 2 defense models. Extensive results demonstrate the superiority of TASAR. Our benchmark enables easy comparisons for future studies, with the code available in the supplementary material.

Vulnerabilities in AI-generated Image Detection: The Challenge of Adversarial Attacks

Recent advancements in image synthesis, particularly with the advent of GAN and Diffusion models, have amplified public concerns regarding the dissemination of disinformation. To address such concerns, numerous AI-generated Image (AIGI) Detectors have been proposed and achieved promising performance in identifying fake images. However, there still lacks a systematic understanding of the adversarial robustness of these AIGI detectors. In this paper, we examine the vulnerability of state-of-the-art AIGI detectors against adversarial attack under white-box and black-box settings, which has been rarely investigated so far. For the task of AIGI detection, we propose a new attack containing two main parts. First, inspired by the obvious difference between real images and fake images in the frequency domain, we add perturbations under the frequency domain to push the image away from its original frequency distribution. Second, we explore the full posterior distribution of the surrogate model to further narrow this gap between heterogeneous models, e.g. transferring adversarial examples across CNNs and ViTs. This is achieved by introducing a novel post-train Bayesian strategy that turns a single surrogate into a Bayesian one, capable of simulating diverse victim models using one pre-trained surrogate, without the need for re-training. We name our method as frequency-based post-train Bayesian attack, or FPBA. Through FPBA, we show that adversarial attack is truly a real threat to AIGI detectors, because FPBA can deliver successful black-box attacks across models, generators, defense methods, and even evade cross-generator detection, which is a crucial real-world detection scenario.

Boosting Adversarial Transferability for Skeleton-based Action Recognition via Exploring the Model Posterior Space

Skeletal motion plays a pivotal role in human activity recognition (HAR). Recently, attack methods have been proposed to identify the universal vulnerability of skeleton-based HAR(S-HAR). However, the research of adversarial transferability on S-HAR is largely missing. More importantly, existing attacks all struggle in transfer across unknown S-HAR models. We observed that the key reason is that the loss landscape of the action recognizers is rugged and sharp. Given the established correlation in prior studies~\cite{qin2022boosting,wu2020towards} between loss landscape and adversarial transferability, we assume and empirically validate that smoothing the loss landscape could potentially improve adversarial transferability on S-HAR. This is achieved by proposing a new post-train Dual Bayesian strategy, which can effectively explore the model posterior space for a collection of surrogates without the need for re-training. Furthermore, to craft adversarial examples along the motion manifold, we incorporate the attack gradient with information of the motion dynamics in a Bayesian manner. Evaluated on benchmark datasets, e.g. HDM05 and NTU 60, the average transfer success rate can reach as high as 35.9\% and 45.5\% respectively. In comparison, current state-of-the-art skeletal attacks achieve only 3.6\% and 9.8\%. The high adversarial transferability remains consistent across various surrogate, victim, and even defense models. Through a comprehensive analysis of the results, we provide insights on what surrogates are more likely to exhibit transferability, to shed light on future research.

Defending Black-box Classifiers by Bayesian Boundary Correction

Classifiers based on deep neural networks have been recently challenged by Adversarial Attack, where the widely existing vulnerability has invoked the research in defending them from potential threats. Given a vulnerable classifier, existing defense methods are mostly white-box and often require re-training the victim under modified loss functions/training regimes. While the model/data/training specifics of the victim are usually unavailable to the user, re-training is unappealing, if not impossible for reasons such as limited computational resources. To this end, we propose a new black-box defense framework. It can turn any pre-trained classifier into a resilient one with little knowledge of the model specifics. This is achieved by new joint Bayesian treatments on the clean data, the adversarial examples and the classifier, for maximizing their joint probability. It is further equipped with a new post-train strategy which keeps the victim intact. We name our framework Bayesian Boundary Correction (BBC). BBC is a general and flexible framework that can easily adapt to different data types. We instantiate BBC for image classification and skeleton-based human activity recognition, for both static and dynamic data. Exhaustive evaluation shows that BBC has superior robustness and can enhance robustness without severely hurting the clean accuracy, compared with existing defense methods.

Unlearnable Examples Give a False Sense of Security: Piercing through Unexploitable Data with Learnable Examples

Safeguarding data from unauthorized exploitation is vital for privacy and security, especially in recent rampant research in security breach such as adversarial/membership attacks. To this end, \textit{unlearnable examples} (UEs) have been recently proposed as a compelling protection, by adding imperceptible perturbation to data so that models trained on them cannot classify them accurately on original clean distribution. Unfortunately, we find UEs provide a false sense of security, because they cannot stop unauthorized users from utilizing other unprotected data to remove the protection, by turning unlearnable data into learnable again. Motivated by this observation, we formally define a new threat by introducing \textit{learnable unauthorized examples} (LEs) which are UEs with their protection removed. The core of this approach is a novel purification process that projects UEs onto the manifold of LEs. This is realized by a new joint-conditional diffusion model which denoises UEs conditioned on the pixel and perceptual similarity between UEs and LEs. Extensive experiments demonstrate that LE delivers state-of-the-art countering performance against both supervised UEs and unsupervised UEs in various scenarios, which is the first generalizable countermeasure to UEs across supervised learning and unsupervised learning.

Understanding the Vulnerability of Skeleton-based Human Activity Recognition via Black-box Attack

Human Activity Recognition (HAR) has been employed in a wide range of applications, e.g. self-driving cars, where safety and lives are at stake. Recently, the robustness of existing skeleton-based HAR methods has been questioned due to their vulnerability to adversarial attacks, which causes concerns considering the scale of the implication. However, the proposed attacks require the full-knowledge of the attacked classifier, which is overly restrictive. In this paper, we show such threats indeed exist, even when the attacker only has access to the input/output of the model. To this end, we propose the very first black-box adversarial attack approach in skeleton-based HAR called BASAR. BASAR explores the interplay between the classification boundary and the natural motion manifold. To our best knowledge, this is the first time data manifold is introduced in adversarial attacks on time series. Via BASAR, we find on-manifold adversarial samples are extremely deceitful and rather common in skeletal motions, in contrast to the common belief that adversarial samples only exist off-manifold. Through exhaustive evaluation, we show that BASAR can deliver successful attacks across classifiers, datasets, and attack modes. By attack, BASAR helps identify the potential causes of the model vulnerability and provides insights on possible improvements. Finally, to mitigate the newly identified threat, we propose a new adversarial training approach by leveraging the sophisticated distributions of on/off-manifold adversarial samples, called mixed manifold-based adversarial training (MMAT). MMAT can successfully help defend against adversarial attacks without compromising classification accuracy.

Defending Black-box Skeleton-based Human Activity Classifiers

Deep learning has been regarded as the `go to' solution for many tasks today, but its intrinsic vulnerability to malicious attacks has become a major concern. The vulnerability is affected by a variety of factors including models, tasks, data, and attackers. Consequently, methods such as Adversarial Training and Randomized Smoothing have been proposed to tackle the problem in a wide range of applications. In this paper, we investigate skeleton-based Human Activity Recognition, which is an important type of time-series data but under-explored in defense against attacks. Our method is featured by (1) a new Bayesian Energy-based formulation of robust discriminative classifiers, (2) a new parameterization of the adversarial sample manifold of actions, and (3) a new post-train Bayesian treatment on both the adversarial samples and the classifier. We name our framework Bayesian Energy-based Adversarial Training or BEAT. BEAT is straightforward but elegant, which turns vulnerable black-box classifiers into robust ones without sacrificing accuracy. It demonstrates surprising and universal effectiveness across a wide range of action classifiers and datasets, under various attacks.

BASAR:Black-box Attack on Skeletal Action Recognition

Skeletal motion plays a vital role in human activity recognition as either an independent data source or a complement. The robustness of skeleton-based activity recognizers has been questioned recently, which shows that they are vulnerable to adversarial attacks when the full-knowledge of the recognizer is accessible to the attacker. However, this white-box requirement is overly restrictive in most scenarios and the attack is not truly threatening. In this paper, we show that such threats do exist under black-box settings too. To this end, we propose the first black-box adversarial attack method BASAR. Through BASAR, we show that adversarial attack is not only truly a threat but also can be extremely deceitful, because on-manifold adversarial samples are rather common in skeletal motions, in contrast to the common belief that adversarial samples only exist off-manifold. Through exhaustive evaluation and comparison, we show that BASAR can deliver successful attacks across models, data, and attack modes. Through harsh perceptual studies, we show that it achieves effective yet imperceptible attacks. By analyzing the attack on different activity recognizers, BASAR helps identify the potential causes of their vulnerability and provides insights on what classifiers are likely to be more robust against attack.