Large language models (LLMs) are becoming increasingly prevalent in modern software systems, interfacing between the user and the Internet to assist with tasks that require advanced language understanding. To accomplish these tasks, the LLM often uses external data sources such as user documents, web retrieval, results from API calls, etc. This opens up new avenues for attackers to manipulate the LLM via prompt injection. Adversarial prompts can be injected into external data sources to override the system's intended instruction and instead execute a malicious instruction. To mitigate this vulnerability, we propose a new defense called SecAlign based on the technique of preference optimization. Our defense first constructs a preference dataset with prompt-injected inputs, secure outputs (ones that respond to the legitimate instruction), and insecure outputs (ones that respond to the injection). We then perform preference optimization on this dataset to teach the LLM to prefer the secure output over the insecure one. This provides the first known method that reduces the success rates of various prompt injections to around 0%, even against attacks much more sophisticated than ones seen during training. This indicates our defense generalizes well against unknown and yet-to-come attacks. Also, our defended models are still practical with similar utility to the one before our defensive training. Our code is at https://github.com/facebookresearch/SecAlign

Large Language Models (LLMs) are attracting significant research attention due to their instruction-following abilities, allowing users and developers to leverage LLMs for a variety of tasks. However, LLMs are vulnerable to prompt-injection attacks: a class of attacks that hijack the model's instruction-following abilities, changing responses to prompts to undesired, possibly malicious ones. In this work, we introduce Jatmo, a method for generating task-specific models resilient to prompt-injection attacks. Jatmo leverages the fact that LLMs can only follow instructions once they have undergone instruction tuning. It harnesses a teacher instruction-tuned model to generate a task-specific dataset, which is then used to fine-tune a base model (i.e., a non-instruction-tuned model). Jatmo only needs a task prompt and a dataset of inputs for the task: it uses the teacher model to generate outputs. For situations with no pre-existing datasets, Jatmo can use a single example, or in some cases none at all, to produce a fully synthetic dataset. Our experiments on seven tasks show that Jatmo models provide similar quality of outputs on their specific task as standard LLMs, while being resilient to prompt injections. The best attacks succeeded in less than 0.5% of cases against our models, versus 87% success rate against GPT-3.5-Turbo. We release Jatmo at https://github.com/wagner-group/prompt-injection-defense.

As Large Language Models (LLMs) are deployed with increasing real-world responsibilities, it is important to be able to specify and constrain the behavior of these systems in a reliable manner. Model developers may wish to set explicit rules for the model, such as "do not generate abusive content", but these may be circumvented by jailbreaking techniques. Evaluating how well LLMs follow developer-provided rules in the face of adversarial inputs typically requires manual review, which slows down monitoring and methods development. To address this issue, we propose Rule-following Language Evaluation Scenarios (RuLES), a programmatic framework for measuring rule-following ability in LLMs. RuLES consists of 15 simple text scenarios in which the model is instructed to obey a set of rules in natural language while interacting with the human user. Each scenario has a concise evaluation program to determine whether the model has broken any rules in a conversation. Through manual exploration of model behavior in our scenarios, we identify 6 categories of attack strategies and collect two suites of test cases: one consisting of unique conversations from manual testing and one that systematically implements strategies from the 6 categories. Across various popular proprietary and open models such as GPT-4 and Llama 2, we find that all models are susceptible to a wide variety of adversarial hand-crafted user inputs, though GPT-4 is the best-performing model. Additionally, we evaluate open models under gradient-based attacks and find significant vulnerabilities. We propose RuLES as a challenging new setting for research into exploring and defending against both manual and automatic attacks on LLMs.

The wide application of deep neural networks (DNNs) demands an increasing amount of attention to their real-world robustness, i.e., whether a DNN resists black-box adversarial attacks, among which score-based query attacks (SQAs) are most threatening since they can effectively hurt a victim network with the only access to model outputs. Defending against SQAs requires a slight but artful variation of outputs due to the service purpose for users, who share the same output information with SQAs. In this paper, we propose a real-world defense by Unifying Gradients (UniG) of different data so that SQAs could only probe a much weaker attack direction that is similar for different samples. Since such universal attack perturbations have been validated as less aggressive than the input-specific perturbations, UniG protects real-world DNNs by indicating attackers a twisted and less informative attack direction. We implement UniG efficiently by a Hadamard product module which is plug-and-play. According to extensive experiments on 5 SQAs, 2 adaptive attacks and 7 defense baselines, UniG significantly improves real-world robustness without hurting clean accuracy on CIFAR10 and ImageNet. For instance, UniG maintains a model of 77.80% accuracy under 2500-query Square attack while the state-of-the-art adversarially-trained model only has 67.34% on CIFAR10. Simultaneously, UniG outperforms all compared baselines in terms of clean accuracy and achieves the smallest modification of the model output. The code is released at https://github.com/snowien/UniG-pytorch.

As data becomes increasingly vital, a company would be very cautious about releasing data, because the competitors could use it to train high-performance models, thereby posing a tremendous threat to the company's commercial competence. To prevent training good models on the data, we could add imperceptible perturbations to it. Since such perturbations aim at hurting the entire training process, they should reflect the vulnerability of DNN training, rather than that of a single model. Based on this new idea, we seek perturbed examples that are always unrecognized (never correctly classified) in training. In this paper, we uncover them by model checkpoints' gradients, forming the proposed self-ensemble protection (SEP), which is very effective because (1) learning on examples ignored during normal training tends to yield DNNs ignoring normal examples; (2) checkpoints' cross-model gradients are close to orthogonal, meaning that they are as diverse as DNNs with different architectures. That is, our amazing performance of ensemble only requires the computation of training one model. By extensive experiments with 9 baselines on 3 datasets and 5 architectures, SEP is verified to be a new state-of-the-art, e.g., our small $\ell_\infty=2/255$ perturbations reduce the accuracy of a CIFAR-10 ResNet18 from 94.56% to 14.68%, compared to 41.35% by the best-known method. Code is available at https://github.com/Sizhe-Chen/SEP.

Although fast adversarial training provides an efficient approach for building robust networks, it may suffer from a serious problem known as catastrophic overfitting (CO), where multi-step robust accuracy suddenly collapses to zero. In this paper, we for the first time decouple single-step adversarial examples into data-information and self-information, which reveals an interesting phenomenon called "self-fitting". Self-fitting, i.e., the network learns the self-information embedded in single-step perturbations, naturally leads to the occurrence of CO. When self-fitting occurs, the network experiences an obvious "channel differentiation" phenomenon that some convolution channels accounting for recognizing self-information become dominant, while others for data-information are suppressed. In this way, the network can only recognize images with sufficient self-information and loses generalization ability to other types of data. Based on self-fitting, we provide new insights into the existing methods to mitigate CO and extend CO to multi-step adversarial training. Our findings reveal a self-learning mechanism in adversarial training and open up new perspectives for suppressing different kinds of information to mitigate CO.

Deep Neural Networks (DNNs) are acknowledged as vulnerable to adversarial attacks, while the existing black-box attacks require extensive queries on the victim DNN to achieve high success rates. For query-efficiency, surrogate models of the victim are used to generate transferable Adversarial Examples (AEs) because of their Gradient Similarity (GS), i.e., surrogates' attack gradients are similar to the victim's ones. However, it is generally neglected to exploit their similarity on outputs, namely the Prediction Similarity (PS), to filter out inefficient queries by surrogates without querying the victim. To jointly utilize and also optimize surrogates' GS and PS, we develop QueryNet, a unified attack framework that can significantly reduce queries. QueryNet creatively attacks by multi-identity surrogates, i.e., crafts several AEs for one sample by different surrogates, and also uses surrogates to decide on the most promising AE for the query. After that, the victim's query feedback is accumulated to optimize not only surrogates' parameters but also their architectures, enhancing both the GS and the PS. Although QueryNet has no access to pre-trained surrogates' prior, it reduces queries by averagely about an order of magnitude compared to alternatives within an acceptable time, according to our comprehensive experiments: 11 victims (including two commercial models) on MNIST/CIFAR10/ImageNet, allowing only 8-bit image queries, and no access to the victim's training data. The code is available at https://github.com/Sizhe-Chen/QueryNet.

Deep neural networks could be fooled by adversarial examples with trivial differences to original samples. To keep the difference imperceptible in human eyes, researchers bound the adversarial perturbations by the $\ell_\infty$ norm, which is now commonly served as the standard to align the strength of different attacks for a fair comparison. However, we propose that using the $\ell_\infty$ norm alone is not sufficient in measuring the attack strength, because even with a fixed $\ell_\infty$ distance, the $\ell_2$ distance also greatly affects the attack transferability between models. Through the discovery, we reach more in-depth understandings towards the attack mechanism, i.e., several existing methods attack black-box models better partly because they craft perturbations with 70% to 130% larger $\ell_2$ distances. Since larger perturbations naturally lead to better transferability, we thereby advocate that the strength of attacks should be simultaneously measured by both the $\ell_\infty$ and $\ell_2$ norm. Our proposal is firmly supported by extensive experiments on ImageNet dataset from 7 attacks, 4 white-box models, and 9 black-box models.

Unlearnable examples (ULEs) aim to protect data from unauthorized usage for training DNNs. Such perturbations, however, are easy to eliminate by adversarial training and data augmentations. In this paper, we resolve this problem from a novel perspective by perturbing only one pixel in each image. Moreover, our produced One-Pixel Shortcut (OPS) could not be erased by adversarial training and strong augmentations. To generate OPS, we perturb in-class images at the same position to the same target value that could mostly and stably deviate from all the original images. Since such generation is only based on images, OPS needs significantly less computational cost than the previous methods using DNN generators. Based on OPS, we introduce an unlearnable dataset called CIFAR-10-S, which is indistinguishable from CIFAR-10 by humans but induces the trained model to extremely low accuracy. Even under adversarial training, a ResNet-18 trained on CIFAR-10-S has only 10.61% accuracy, compared to 83.02% by the existing error-minimizing method. Deep neural networks (DNNs) have successfully promoted the computer vision field in the past decade. As DNNs are scaling up unprecedentedly (Brock et al., 2018; Huang et al., 2019; Riquelme et al., 2021; Zhang et al., 2022), data becomes increasingly vital. For example, ImageNet (Russakovsky et al., 2015) fostered the development of AlexNet (Krizhevsky et al., 2017).

The score-based query attacks (SQAs) pose practical threats to deep neural networks by crafting adversarial perturbations within dozens of queries, only using the model's output scores. Nonetheless, we note that if the loss trend of the outputs is slightly perturbed, SQAs could be easily misled and thereby become much less effective. Following this idea, we propose a novel defense, namely Adversarial Attack on Attackers (AAA), to confound SQAs towards incorrect attack directions by slightly modifying the output logits. In this way, (1) SQAs are prevented regardless of the model's worst-case robustness; (2) the original model predictions are hardly changed, i.e., no degradation on clean accuracy; (3) the calibration of confidence scores can be improved simultaneously. Extensive experiments are provided to verify the above advantages. For example, by setting $\ell_\infty=8/255$ on CIFAR-10, our proposed AAA helps WideResNet-28 secure 80.59% accuracy under Square attack (2500 queries), while the best prior defense (i.e., adversarial training) only attains 67.44%. Since AAA attacks SQA's general greedy strategy, such advantages of AAA over 8 defenses can be consistently observed on 8 CIFAR-10/ImageNet models under 6 SQAs, using different attack targets, bounds, norms, losses, and strategies. Moreover, AAA calibrates better without hurting the accuracy. Our code is available at https://github.com/Sizhe-Chen/AAA.