NeurIPS 2023 Papers

We study indiscriminate poisoning for linear learners where an adversary injects a few crafted examples into the training data with the goal of forcing the induced model to incur higher test error. Inspired by the observation that linear learners on some datasets are able to resist the best known attacks even without any defenses, we further investigate whether datasets can be inherently robust to indiscriminate poisoning attacks for linear learners. For theoretical Gaussian distributions, we rigorously characterize the behavior of an optimal poisoning attack, defined as the poisoning strategy that attains the maximum risk of the induced model at a given poisoning budget. Our results prove that linear learners can indeed be robust to indiscriminate poisoning if the class-wise data distributions are well-separated with low variance and the size of the constraint set containing all permissible poisoning points is also small. These findings largely explain the drastic variation in empirical attack performance of the state-of-the-art poisoning attacks on linear learners across benchmark datasets, making an important initial step towards understanding the underlying reasons some learning tasks are vulnerable to data poisoning attacks.

Diffusion Models (DMs) are state-of-the-art generative models that learn a reversible corruption process from iterative noise addition and denoising. They are the backbone of many generative AI applications, such as text-to-image conditional generation. However, recent studies have shown that basic unconditional DMs (e.g., DDPM and DDIM) are vulnerable to backdoor injection, a type of output manipulation attack triggered by a maliciously embedded pattern at model input. This paper presents a unified backdoor attack framework (VillanDiffusion) to expand the current scope of backdoor analysis for DMs. Our framework covers mainstream unconditional and conditional DMs (denoising-based and score-based) and various training-free samplers for holistic evaluations. Experiments show that our unified framework facilitates the backdoor analysis of different DM configurations and provides new insights into caption-based backdoor attacks on DMs.

With the growing demand for the right to be forgotten, there is an increasing need for machine learning models to forget sensitive data and its impact. To address this, the paradigm of selective forgetting (a.k.a machine unlearning) has been extensively studied, which aims to remove the impact of requested data from a well-trained model without retraining from scratch. Despite its significant success, limited attention has been given to the security vulnerabilities of the unlearning system concerning malicious data update requests. Motivated by this, in this paper, we explore the possibility and feasibility of malicious data update requests during the unlearning process. Specifically, we first propose a new class of malicious selective forgetting attacks, which involves a static scenario where all the malicious data update requests are provided by the adversary at once. Additionally, considering the sequential setting where the data update requests arrive sequentially, we also design a novel framework for sequential forgetting attacks, which is formulated as a stochastic optimal control problem. We also propose novel optimization algorithms that can find the effective malicious data update requests. We perform theoretical analyses for the proposed selective forgetting attacks, and extensive experimental results validate the effectiveness of our proposed selective forgetting attacks. The source code is available in the supplementary material.

Federated learning algorithms enable neural network models to be trained across multiple decentralized edge devices without sharing private data. However, they are susceptible to backdoor attacks launched by malicious clients. Existing robust federated aggregation algorithms heuristically detect and exclude suspicious clients based on their parameter distances, but they are ineffective on Natural Language Processing (NLP) tasks. The main reason is that, although text backdoor patterns are obvious at the underlying dataset level, they are usually hidden at the parameter level, since injecting backdoors into texts with discrete feature space has less impact on the statistics of the model parameters. To settle this issue, we propose to identify backdoor clients by explicitly modeling the data divergence among clients in federated NLP systems. Through theoretical analysis, we derive the f-divergence indicator to estimate the client data divergence with aggregation updates and Hessians. Furthermore, we devise a dataset synthesization method with a Hessian reassignment mechanism guided by the diffusion theory to address the key challenge of inaccessible datasets in calculating clients' data Hessians. We then present the novel Federated F-Divergence-Based Aggregation~(\textbf{Fed-FA}) algorithm, which leverages the f-divergence indicator to detect and discard suspicious clients. Extensive empirical results show that Fed-FA outperforms all the parameter distance-based methods in defending against backdoor attacks among various natural language backdoor attack scenarios.

The adversarial vulnerability of Deep Neural Networks (DNNs) has been well-known and widely concerned, often under the context of learning top-$1$ attacks (e.g., fooling a DNN to classify a cat image as dog). This paper shows that the concern is much more serious by learning significantly more aggressive ordered top-$K$ clear-box targeted attacks proposed in~\citep{zhang2020learning}. We propose a novel and rigorous quadratic programming (QP) method of learning ordered top-$K$ attacks with low computing cost, dubbed as \textbf{QuadAttac$K$}. Our QuadAttac$K$ directly solves the QP to satisfy the attack constraint in the feature embedding space (i.e., the input space to the final linear classifier), which thus exploits the semantics of the feature embedding space (i.e., the principle of class coherence). With the optimized feature embedding vector perturbation, it then computes the adversarial perturbation in the data space via the vanilla one-step back-propagation. In experiments, the proposed QuadAttac$K$ is tested in the ImageNet-1k classification using ResNet-50, DenseNet-121, and Vision Transformers (ViT-B and DEiT-S). It successfully pushes the boundary of successful ordered top-$K$ attacks from $K=10$ up to $K=20$ at a cheap budget ($1\times 60$) and further improves attack success rates for $K=5$ for all tested models, while retaining the performance for $K=1$.

Membership inference attacks are designed to determine, using black box access to trained models, whether a particular example was used in training or not. Membership inference can be formalized as a hypothesis testing problem. The most effective existing attacks estimate the distribution of some test statistic (usually the model's confidence on the true label) on points that were (and were not) used in training by training many \emph{shadow models}---i.e. models of the same architecture as the model being attacked, trained on a random subsample of data. While effective, these attacks are extremely computationally expensive, especially when the model under attack is large. \footnotetext[0]{ Martin and Shuai are the lead authors, and other authors are ordered alphabetically. \{maberlop,shuat\}@amazon.com} We introduce a new class of attacks based on performing quantile regression on the distribution of confidence scores induced by the model under attack on points that are not used in training. We show that our method is competitive with state-of-the-art shadow model attacks, while requiring substantially less compute because our attack requires training only a single model. Moreover, unlike shadow model attacks, our proposed attack does not require any knowledge of the architecture of the model under attack and is therefore truly ``black-box". We show the efficacy of this approach in an extensive series of experiments on various datasets and model architectures. Our code is available at \href{https://github.com/amazon-science/quantile-mia}{github.com/amazon-science/quantile-mia.}

In a backdoor attack, an adversary injects corrupted data into a model's training dataset in order to gain control over its predictions on images with a specific attacker-defined trigger. A typical corrupted training example requires altering both the image, by applying the trigger, and the label. Models trained on clean images, therefore, were considered safe from backdoor attacks. However, in some common machine learning scenarios, the training labels are provided by potentially malicious third-parties. This includes crowd-sourced annotation and knowledge distillation. We, hence, investigate a fundamental question: can we launch a successful backdoor attack by only corrupting labels? We introduce a novel approach to design label-only backdoor attacks, which we call FLIP, and demonstrate its strengths on three datasets (CIFAR-10, CIFAR-100, and Tiny-ImageNet) and four architectures (ResNet-32, ResNet-18, VGG-19, and Vision Transformer). With only 2% of CIFAR-10 labels corrupted, FLIP achieves a near-perfect attack success rate of 99.4% while suffering only a 1.8% drop in the clean test accuracy. Our approach builds upon the recent advances in trajectory matching, originally introduced for dataset distillation.

Federated learning (FL) provides a distributed training paradigm where multiple clients can jointly train a global model without sharing their local data. However, recent studies have shown that FL offers an additional surface for backdoor attacks. For instance, an attacker can compromise a subset of clients and thus corrupt the global model to misclassify an input with a backdoor trigger as the adversarial target. Existing defenses for FL against backdoor attacks usually detect and exclude the corrupted information from the compromised clients based on a static attacker model. However, such defenses are inadequate against dynamic attackers who strategically adapt their attack strategies. To bridge this gap, we model the strategic interactions between the defender and dynamic attackers as a minimax game. Based on the analysis of the game, we design an interactive defense mechanism FedGame. We prove that under mild assumptions, the global model trained with FedGame under backdoor attacks is close to that trained without attacks. Empirically, we compare FedGame with multiple state-of-the-art baselines on several benchmark datasets under various attacks. We show that FedGame can effectively defend against strategic attackers and achieves significantly higher robustness than baselines. Our code is available at: https://github.com/AI-secure/FedGame.

Backdoor attacks pose a severe threat to the supply chain management of deep reinforcement learning (DRL) policies. Despite initial defenses proposed in recent studies, these methods have very limited generalizability and scalability. To address this issue, we propose BIRD, a technique to detect and remove backdoors from a pretrained DRL policy in a clean environment without requiring any knowledge about the attack specifications and accessing its training process. By analyzing the unique properties and behaviors of backdoor attacks, we formulate trigger restoration as an optimization problem and design a novel metric to detect backdoored policies. We also design a finetuning method to remove the backdoor, while maintaining the agent's performance in the clean environment. We evaluate BIRD against three backdoor attacks in ten different single-agent or multi-agent environments. Our results verify the effectiveness, efficiency, and generalizability of BIRD, as well as its robustness to different attack variations and adaptions.

Federated learning (FL) is a distributed learning approach that enables machine learning models to be trained on decentralized data without compromising end devices' personal, potentially sensitive data. However, the distributed nature and uninvestigated data intuitively introduce new security vulnerabilities, including backdoor attacks. In this scenario, an adversary implants backdoor functionality into the global model during training, which can be activated to cause the desired misbehaviors for any input with a specific adversarial pattern. Despite having remarkable success in triggering and distorting model behavior, prior backdoor attacks in FL often hold impractical assumptions, limited imperceptibility, and durability. Specifically, the adversary needs to control a sufficiently large fraction of clients or know the data distribution of other honest clients. In many cases, the trigger inserted is often visually apparent, and the backdoor effect is quickly diluted if the adversary is removed from the training process. To address these limitations, we propose a novel backdoor attack framework in FL, the Irreversible Backdoor Attack (IBA), that jointly learns the optimal and visually stealthy trigger and then gradually implants the backdoor into a global model. This approach allows the adversary to execute a backdoor attack that can evade both human and machine inspections. Additionally, we enhance the efficiency and durability of the proposed attack by selectively poisoning the model's parameters that are least likely updated by the main task's learning process and constraining the poisoned model update to the vicinity of the global model. Finally, we evaluate the proposed attack framework on several benchmark datasets, including MNIST, CIFAR-10, and Tiny ImageNet, and achieved high success rates while simultaneously bypassing existing backdoor defenses and achieving a more durable backdoor effect compared to other backdoor attacks. Overall, IBA offers a more effective, stealthy, and durable approach to backdoor attacks in FL. The code associated with this paper is available on [GitHub](https://github.com/sail-research/iba).

Federated Learning (FL) is a distributed machine learning paradigm that allows multiple clients to train a global model collaboratively without sharing their local training data. Due to its distributed nature, many studies have shown that it is vulnerable to backdoor attacks. However, existing studies usually used a predetermined, fixed backdoor trigger or optimized it based solely on the local data and model without considering the global training dynamics. This leads to sub-optimal and less durable attack effectiveness, i.e., their attack success rate is low when the attack budget is limited and decreases quickly if the attacker can no longer perform attacks anymore. To address these limitations, we propose A3FL, a new backdoor attack which adversarially adapts the backdoor trigger to make it less likely to be removed by the global training dynamics. Our key intuition is that the difference between the global model and the local model in FL makes the local-optimized trigger much less effective when transferred to the global model. We solve this by optimizing the trigger to even survive the worst-case scenario where the global model was trained to directly unlearn the trigger. Extensive experiments on benchmark datasets are conducted for twelve existing defenses to comprehensively evaluate the effectiveness of our A3FL. Our code is available at https://github.com/hfzhang31/A3FL.

Contrastive vision-language representation learning has achieved state-of-the-art performance for zero-shot classification, by learning from millions of image-caption pairs crawled from the internet. However, the massive data that powers large multimodal models such as CLIP, makes them extremely vulnerable to various types of targeted data poisoning and backdoor attacks. Despite this vulnerability, robust contrastive vision-language pre-training against such attacks has remained unaddressed. In this work, we propose RoCLIP, the first effective method for robust pre-training multimodal vision-language models against targeted data poisoning and backdoor attacks. RoCLIP effectively breaks the association between poisoned image-caption pairs by considering a relatively large and varying pool of random captions, and matching every image with the text that is most similar to it in the pool instead of its own caption, every few epochs.It also leverages image and text augmentations to further strengthen the defense and improve the performance of the model. Our extensive experiments show that RoCLIP renders state-of-the-art targeted data poisoning and backdoor attacks ineffective during pre-training CLIP models. In particular, RoCLIP decreases the success rate for targeted data poisoning attacks from 93.75% to 12.5% and that of backdoor attacks down to 0%, while improving the model's linear probe performance by 10% and maintains a similar zero shot performance compared to CLIP. By increasing the frequency of matching, RoCLIP is able to defend strong attacks, which add up to 1% poisoned examples to the data, and successfully maintain a low attack success rate of 12.5%, while trading off the performance on some tasks.

We introduce camouflaged data poisoning attacks, a new attack vector that arises in the context of machine unlearning and other settings when model retraining may be induced. An adversary first adds a few carefully crafted points to the training dataset such that the impact on the model's predictions is minimal. The adversary subsequently triggers a request to remove a subset of the introduced points at which point the attack is unleashed and the model's predictions are negatively affected. In particular, we consider clean-label targeted attacks (in which the goal is to cause the model to misclassify a specific test point) on datasets including CIFAR-10, Imagenette, and Imagewoof. This attack is realized by constructing camouflage datapoints that mask the effect of a poisoned dataset. We demonstrate efficacy of our attack when unlearning is performed via retraining from scratch, the idealized setting of machine unlearning which other efficient methods attempt to emulate, as well as against the approximate unlearning approach of Graves et al. (2021).

Pre-trained language models (PLMs) have demonstrated remarkable performance as few-shot learners. However, their security risks under such settings are largely unexplored. In this work, we conduct a pilot study showing that PLMs as few-shot learners are highly vulnerable to backdoor attacks while existing defenses are inadequate due to the unique challenges of few-shot scenarios. To address such challenges, we advocate MDP, a novel lightweight, pluggable, and effective defense for PLMs as few-shot learners. Specifically, MDP leverages the gap between the masking-sensitivity of poisoned and clean samples: with reference to the limited few-shot data as distributional anchors, it compares the representations of given samples under varying masking and identifies poisoned samples as ones with significant variations. We show analytically that MDP creates an interesting dilemma for the attacker to choose between attack effectiveness and detection evasiveness. The empirical evaluation using benchmark datasets and representative attacks validates the efficacy of MDP. The code of MDP is publicly available.

Backdoor attacks have emerged as a prominent threat to natural language processing (NLP) models, where the presence of specific triggers in the input can lead poisoned models to misclassify these inputs to predetermined target classes. Current detection mechanisms are limited by their inability to address more covert backdoor strategies, such as style-based attacks. In this work, we propose an innovative test-time poisoned sample detection framework that hinges on the interpretability of model predictions, grounded in the semantic meaning of inputs. We contend that triggers (e.g., infrequent words) are not supposed to fundamentally alter the underlying semantic meanings of poisoned samples as they want to stay stealthy. Based on this observation, we hypothesize that while the model's predictions for paraphrased clean samples should remain stable, predictions for poisoned samples should revert to their true labels upon the mutations applied to triggers during the paraphrasing process. We employ ChatGPT, a state-of-the-art large language model, as our paraphraser and formulate the trigger-removal task as a prompt engineering problem. We adopt fuzzing, a technique commonly used for unearthing software vulnerabilities, to discover optimal paraphrase prompts that can effectively eliminate triggers while concurrently maintaining input semantics. Experiments on 4 types of backdoor attacks, including the subtle style backdoors, and 4 distinct datasets demonstrate that our approach surpasses baseline methods, including STRIP, RAP, and ONION, in precision and recall.

Object detection models are vulnerable to backdoor or trojan attacks, where an attacker can inject malicious triggers into the model, leading to altered behavior during inference. As a defense mechanism, trigger inversion leverages optimization to reverse-engineer triggers and identify compromised models. While existing trigger inversion methods assume that each instance from the support set is equally affected by the injected trigger, we observe that the poison effect can vary significantly across bounding boxes in object detection models due to its dense prediction nature, leading to an undesired optimization objective misalignment issue for existing trigger reverse-engineering methods. To address this challenge, we propose the first object detection backdoor detection framework Django (Detecting Trojans in Object Detection Models via Gaussian Focus Calibration). It leverages a dynamic Gaussian weighting scheme that prioritizes more vulnerable victim boxes and assigns appropriate coefficients to calibrate the optimization objective during trigger inversion. In addition, we combine Django with a novel label proposal pre-processing technique to enhance its efficiency. We evaluate Django on 3 object detection image datasets, 3 model architectures, and 2 types of attacks, with a total of 168 models. Our experimental results show that Django outperforms 6 state-of-the-art baselines, with up to 38% accuracy improvement and 10x reduced overhead. The code is available at https://github.com/PurduePAML/DJGO.

It has been widely observed that deep neural networks (DNN) are vulnerable to backdoor attacks where attackers could manipulate the model behavior maliciously by tampering with a small set of training samples. Although a line of defense methods is proposed to mitigate this threat, they either require complicated modifications to the training process or heavily rely on the specific model architecture, which makes them hard to deploy into real-world applications. Therefore, in this paper, we instead start with fine-tuning, one of the most common and easy-to-deploy backdoor defenses, through comprehensive evaluations against diverse attack scenarios. Observations made through initial experiments show that in contrast to the promising defensive results on high poisoning rates, vanilla tuning methods completely fail at low poisoning rate scenarios. Our analysis shows that with the low poisoning rate, the entanglement between backdoor and clean features undermines the effect of tuning-based defenses. Therefore, it is necessary to disentangle the backdoor and clean features in order to improve backdoor purification. To address this, we introduce Feature Shift Tuning (FST), a method for tuning-based backdoor purification. Specifically, FST encourages feature shifts by actively deviating the classifier weights from the originally compromised weights. Extensive experiments demonstrate that our FST provides consistently stable performance under different attack settings. Without complex parameter adjustments, FST also achieves much lower tuning costs, only $10$ epochs. Our codes are available at https://github.com/AISafety-HKUST/stable_backdoor_purification.

Backdoor attacks involve inserting poisoned samples during training, resulting in a model containing a hidden backdoor that can trigger specific behaviors without impacting performance on normal samples. These attacks are challenging to detect, as the backdoored model appears normal until activated by the backdoor trigger, rendering them particularly stealthy. In this study, we devise a unified inference-stage detection framework to defend against backdoor attacks. We first rigorously formulate the inference-stage backdoor detection problem, encompassing various existing methods, and discuss several challenges and limitations. We then propose a framework with provable guarantees on the false positive rate or the probability of misclassifying a clean sample. Further, we derive the most powerful detection rule to maximize the detection power, namely the rate of accurately identifying a backdoor sample, given a false positive rate under classical learning scenarios. Based on the theoretically optimal detection rule, we suggest a practical and effective approach for real-world applications based on the latent representations of backdoored deep nets. We extensively evaluate our method on 14 different backdoor attacks using Computer Vision (CV) and Natural Language Processing (NLP) benchmark datasets. The experimental findings align with our theoretical results. We significantly surpass the state-of-the-art methods, e.g., up to 300\% improvement on the detection power as evaluated by AUCROC, over the state-of-the-art defense against advanced adaptive backdoor attacks.

Model poisoning attacks greatly jeopardize the application of federated learning (FL). The effectiveness of existing defenses is susceptible to the latest model poisoning attacks, leading to a decrease in prediction accuracy. Besides, these defenses are intractable to distinguish benign outliers from malicious gradients, which further compromises the model generalization. In this work, we propose a novel defense including detection and aggregation, named RECESS, to serve as a “vaccine” for FL against model poisoning attacks. Different from the passive analysis in previous defenses, RECESS proactively queries each participating client with a delicately constructed aggregation gradient, accompanied by the detection of malicious clients according to their responses with higher accuracy. Further, RECESS adopts a newly proposed trust scoring based mechanism to robustly aggregate gradients. Rather than previous methods of scoring in each iteration, RECESS takes into account the correlation of clients’ performance over multiple iterations to estimate the trust score, bringing in a significant increase in detection fault tolerance. Finally, we extensively evaluate RECESS on typical model architectures and four datasets under various settings including white/black-box, cross-silo/device FL, etc. Experimental results show the superiority of RECESS in terms of reducing accuracy loss caused by the latest model poisoning attacks over five classic and two state-of-the-art defenses.