ICLR 2024 Papers

While Large Language Models (LLMs) are increasingly being used in real-world applications, they remain vulnerable to *prompt injection attacks*: malicious third party prompts that subvert the intent of the system designer. To help researchers study this problem, we present a dataset of over 563,000 prompt injection attacks and 118,000 prompt-based "defenses" against prompt injection, all created by players of an online game called Tensor Trust. To the best of our knowledge, this is the first dataset that includes both human-generated attacks and defenses for instruction-following LLMs. The attacks in our dataset have easily interpretable structure, and shed light on the weaknesses of LLMs. We also use the dataset to create a benchmark for resistance to two types of prompt injection, which we refer to as *prompt extraction* and *prompt hijacking*. Our benchmark results show that many models are vulnerable to the attack strategies in the Tensor Trust dataset. Furthermore, we show that some attack strategies from the dataset generalize to deployed LLM-based applications, even though they have a very different set of constraints to the game. We release data and code at [tensortrust.ai/paper](https://tensortrust.ai/paper)

In this paper, we revisit the problem of sparse linear regression in the local differential privacy (LDP) model. Existing research in the non-interactive and sequentially local models has focused on obtaining the lower bounds for the case where the underlying parameter is $1$-sparse, and extending such bounds to the more general $k$-sparse case has proven to be challenging. Moreover, it is unclear whether efficient non-interactive LDP (NLDP) algorithms exist. To address these issues, we first consider the problem in the $\epsilon$ non-interactive LDP model and provide a lower bound of $\Omega(\frac{\sqrt{dk\log d}}{\sqrt{n}\epsilon})$ on the $\ell_2$-norm estimation error for sub-Gaussian data, where $n$ is the sample size and $d$ is the dimension of the space. We propose an innovative NLDP algorithm, the very first of its kind for the problem. As a remarkable outcome, this algorithm also yields a novel and highly efficient estimator as a valuable by-product. Our algorithm achieves an upper bound of $\tilde{O}({\frac{d\sqrt{k}}{\sqrt{n}\epsilon}})$ for the estimation error when the data is sub-Gaussian, which can be further improved by a factor of $O(\sqrt{d})$ if the server has additional public but unlabeled data. For the sequentially interactive LDP model, we show a similar lower bound of $\Omega({\frac{\sqrt{dk}}{\sqrt{n}\epsilon}})$. As for the upper bound, we rectify a previous method and show that it is possible to achieve a bound of $\tilde{O}(\frac{k\sqrt{d}}{\sqrt{n}\epsilon})$. Our findings reveal fundamental differences between the non-private case, central DP model, and local DP model in the sparse linear regression problem.

Recent research has highlighted the vulnerability of Deep Neural Networks (DNNs) against data poisoning attacks. These attacks aim to inject poisoning samples into the models' training dataset such that the trained models have inference failures. While previous studies have executed different types of attacks, one major challenge that greatly limits their effectiveness is the uncertainty of the re-training process after the injection of poisoning samples. It includes the uncertainty of training initialization, algorithm and model architecture. To address this challenge, we propose a new strategy called **Sharpness-Aware Data Poisoning Attack (SAPA)**. In particular, it leverages the concept of DNNs' loss landscape sharpness to optimize the poisoning effect on the (approximately) worst re-trained model. Extensive experiments demonstrate that SAPA offers a general and principled strategy that significantly enhances various types of poisoning attacks against various types of re-training uncertainty.

Fair machine learning seeks to mitigate model prediction bias against certain demographic subgroups such as elder and female. Recently, fair representation learning (FRL) trained by deep neural networks has demonstrated superior performance, whereby representations containing no demographic information are inferred from the data and then used as the input to classification or other downstream tasks. Despite the development of FRL methods, their vulnerability under data poisoning attack, a popular protocol to benchmark model robustness under adversarial scenarios, is under-explored. Data poisoning attacks have been developed for classical fair machine learning methods which incorporate fairness constraints into shallow-model classifiers. Nonetheless, these attacks fall short in FRL due to notably different fairness goals and model architectures. This work proposes the first data poisoning framework attacking FRL. We induce the model to output unfair representations that contain as much demographic information as possible by injecting carefully crafted poisoning samples into the training data. This attack entails a prohibitive bilevel optimization, wherefore an effective approximated solution is proposed. A theoretical analysis on the needed number of poisoning samples is derived and sheds light on defending against the attack. Experiments on benchmark fairness datasets and state-of-the-art fair representation learning models demonstrate the superiority of our attack.

We introduce Tangent Attention Fine-Tuning (TAFT), a method for fine-tuning linearized transformers obtained by computing a First-order Taylor Expansion around a pre-trained initialization. We show that the Jacobian-Vector Product resulting from linearization can be computed efficiently in a single forward pass, reducing training and inference cost to the same order of magnitude as its original non-linear counterpart, while using the same number of parameters. Furthermore, we show that, when applied to various downstream visual classification tasks, the resulting Tangent Transformer fine-tuned with TAFT can perform comparably with fine-tuning the original non-linear network. Since Tangent Transformers are linear with respect to the new set of weights, and the resulting fine-tuning loss is convex, we show that TAFT enjoys several advantages compared to non-linear fine-tuning when it comes to model composition, parallel training, machine unlearning, and differential privacy. Our code is available at: https://github.com/tianyu139/tangent-model-composition

Adversarial attacks have been a looming and unaddressed threat in the industry. However, through a decade-long history of the robustness evaluation literature, we have learned that mounting a strong or optimal attack is challenging. It requires both machine learning and domain expertise. In other words, the white-box threat model, religiously assumed by a large majority of the past literature, is unrealistic. In this paper, we propose a new practical threat model where the adversary relies on transfer attacks through publicly available surrogate models. We argue that this setting will become the most prevalent for security-sensitive applications in the future. We evaluate the transfer attacks in this setting and propose a specialized defense method based on a game-theoretic perspective. The defenses are evaluated under 24 public models and 11 attack algorithms across three datasets (CIFAR-10, CIFAR-100, and ImageNet). Under this threat model, our defense, PubDef, outperforms the state-of-the-art white-box adversarial training by a large margin with almost no loss in the normal accuracy. For instance, on ImageNet, our defense achieves 62% accuracy under the strongest transfer attack vs only 36% of the best adversarially trained model. Its accuracy when not under attack is only 2% lower than that of an undefended model (78% vs 80%). We release our code at https://github.com/wagner-group/pubdef.

Post hoc privacy auditing techniques can be used to test the privacy guarantees of a model, but come with several limitations: (i) they can only establish lower bounds on the privacy loss, (ii) the intermediate model updates and some data must be shared with the auditor to get a better approximation of the privacy loss, and (iii) the auditor typically faces a steep computational cost to run a large number of attacks. In this paper, we propose to proactively generate a cryptographic certificate of privacy during training to forego such auditing limitations. We introduce Confidential-DPproof , a framework for Confidential Proof of Differentially Private Training, which enhances training with a certificate of the $(\varepsilon,\delta)$-DP guarantee achieved. To obtain this certificate without revealing information about the training data or model, we design a customized zero-knowledge proof protocol tailored to the requirements introduced by differentially private training, including random noise addition and privacy amplification by subsampling. In experiments on CIFAR-10, Confidential-DPproof trains a model achieving state-of-the-art $91$% test accuracy with a certified privacy guarantee of $(\varepsilon=0.55,\delta=10^{-5})$-DP in approximately 100 hours.

We study the unique, less-well understood problem of generating sparse adversarial samples simply by observing the score-based replies to model queries. Sparse attacks aim to discover a minimum number—the $l_0$ bounded—perturbations to model inputs to craft adversarial examples and misguide model decisions. But, in contrast to query-based dense attack counterparts against black-box models, constructing sparse adversarial perturbations, even when models serve confidence score information to queries in a score-based setting, is non-trivial. Because, such an attack leads to: i) an NP-hard problem; and ii) a non-differentiable search space. We develop the BRUSLEATTACK—a new, faster (more query-efficient) algorithm formulation for the problem. We conduct extensive attack evaluations including an attack demonstration against a Machine Learning as a Service (MLaaS) offering exemplified by __Google Cloud Vision__ and robustness testing of adversarial training regimes and a recent defense against black-box attacks. The proposed attack scales to achieve state-of-the-art attack success rates and query efficiency on standard computer vision tasks such as ImageNet across different model architectures. Our artifacts and DIY attack samples are available on GitHub. Importantly, our work facilitates faster evaluation of model vulnerabilities and raises our vigilance on the safety, security and reliability of deployed systems.

Untrustworthy users can misuse image generators to synthesize high-quality deepfakes and engage in unethical activities. Watermarking deters misuse by marking generated content with a hidden message, enabling its detection using a secret watermarking key. A core security property of watermarking is robustness, which states that an attacker can only evade detection by substantially degrading image quality. Assessing robustness requires designing an adaptive attack for the specific watermarking algorithm. When evaluating watermarking algorithms and their (adaptive) attacks, it is challenging to determine whether an adaptive attack is optimal, i.e., the best possible attack. We solve this problem by defining an objective function and then approach adaptive attacks as an optimization problem. The core idea of our adaptive attacks is to replicate secret watermarking keys locally by creating surrogate keys that are differentiable and can be used to optimize the attack's parameters. We demonstrate for Stable Diffusion models that such an attacker can break all five surveyed watermarking methods at no visible degradation in image quality. Optimizing our attacks is efficient and requires less than 1 GPU hour to reduce the detection accuracy to 6.3% or less. Our findings emphasize the need for more rigorous robustness testing against adaptive, learnable attackers.

Low-rank adaptation (LoRA) is one of the most popular task-specific parameter-efficient fine-tuning (PEFT) methods on pre-trained language models for its good performance and computational efficiency. LoRA injects a product of two trainable rank decomposition matrices over the top of each frozen pre-trained model module. However, when applied in the setting of privacy-preserving federated learning (FL), LoRA may become unstable due to the following facts: 1) the effects of data heterogeneity and multi-step local updates are non-negligible, 2) additive noise enforced on updating gradients to guarantee differential privacy (DP) can be amplified and 3) the final performance is susceptible to hyper-parameters. A key factor leading to these phenomena is the discordance between jointly optimizing the two low-rank matrices by local clients and separately aggregating them by the central server. Thus, this paper proposes an efficient and effective version of LoRA, Federated Freeze A LoRA (FFA-LoRA), to alleviate these challenges and further halve the communication cost of federated fine-tuning LLMs. The core idea of FFA-LoRA is to fix the randomly initialized non-zero matrices and only fine-tune the zero-initialized matrices. Compared to LoRA, FFA-LoRA is motivated by practical and theoretical benefits in privacy-preserved FL. Our experiments demonstrate that FFA-LoRA provides more consistent performance with better computational efficiency over vanilla LoRA in various FL tasks.

An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model, such as a proprietary system like ChatGPT (e.g., Alpaca, Self-Instruct, and others). In this work, we critically analyze this approach of imitating language models. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes (1.5B--13B), data sources, and imitation data amounts (0.3M--150M tokens). We then evaluate the models using crowd raters and canonical NLP benchmarks. Initially, we were surprised by the output quality of our imitation models---they appear far better at following instructions, and crowd workers rate their outputs as competitive with ChatGPT. However, when conducting more targeted automatic evaluations, we find that imitation models close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data. We show that these performance discrepancies may slip past human raters because imitation models are adept at mimicking ChatGPT’s style but not its factuality. Overall, we conclude that while model imitation can be useful for training models to follow instructions and avoid toxic outputs, it falls short its full promise in many ways. In particular, there exists a substantial capabilities gap between open and closed LMs that we find cannot be bridged merely by adding more imitation data. Instead, we find that fine-tuning more capable base LMs has a significantly more substantial effect on closing this gap. In turn, we argue that the higher leverage action for improving open-source models is to tackle the difficult challenge of developing better base LMs, rather than taking the shortcut of imitating proprietary systems.

Label differentially private (label DP) algorithms seek to preserve the privacy of the labels in a training dataset in settings where the features are known to the adversary. In this work, we study a new family of label DP training algorithms. Unlike most prior label DP algorithms that have been based on label randomization, our algorithm naturally leverages the power of the central model of DP. It interleaves gradient projection operations with private stochastic gradient descent steps in order to improve the utility of the trained model while guaranteeing the privacy of the labels. We show that such projection-based algorithms can be made practical and that they improve on the state-of-the art for label DP training in the high-privacy regime. We complement our empirical evaluation with theoretical results shedding light on the efficacy of our method through the lens of bias-variance trade-offs.

Node labels for graphs are usually generated using an automated process or crowd-sourced from human users. This opens up avenues for malicious users to compromise the training labels, making it unwise to blindly rely on them. While robustness against noisy labels is an active area of research, there are only a handful of papers in the literature that address this for graph-based data. Even more so, the effects of adversarial label perturbations is sparsely studied. More critically, we reveal that the entire literature on label poisoning for GNNs is plagued by serious evaluation pitfalls. Thus making it hard to conclude how robust GNNs are against label perturbations. After course correcting the state of label poisoning attacks with our faithful evaluation, we identify a discrepancy in attack efficiency of $\sim9\%$ on average. Additionally, we introduce two new simple yet effective attacks that are significantly stronger (up to $\sim8\%$) than the previous strongest attack. Our strongest proposed attack can be efficiently computed and is theoretically backed.

Large Language Models (LLMs) have emerged as dominant tools for various tasks, particularly when tailored for a specific target by prompt tuning. Nevertheless, concerns surrounding data privacy present obstacles due to the tuned prompts' dependency on sensitive private information. A practical solution is to host a local LLM and optimize a soft prompt privately using data. Yet, hosting a local model becomes problematic when model ownership is protected. Alternative methods, like sending data to the model's provider for training, intensify these privacy issues facing an untrusted provider. In this paper, we present a novel solution called Differentially-Private Offsite Prompt Tuning (DP-OPT) to address this challenge. Our approach involves tuning a discrete prompt on the client side and then applying it to the desired cloud models. We demonstrate that prompts suggested by LLMs themselves can be transferred without compromising performance significantly. To ensure that the prompts do not leak private information, we introduce the first private prompt generation mechanism, by a differentially-private (DP) ensemble of in-context learning with private demonstrations. With DP-OPT, generating privacy-preserving prompts by Vicuna-7b can yield competitive performance compared to non-private in-context learning on GPT3.5 or local private prompt tuning. Codes are available at https://github.com/VITA-Group/DP-OPT.

Graph classification, which aims to predict a label for a graph, has many real-world applications such as malware detection, fraud detection, and healthcare. However, many studies show an attacker could carefully perturb the structure and/or node features in a graph such that a graph classifier misclassifies the perturbed graph. Such vulnerability impedes the deployment of graph classification in security/safety-critical applications. Existing empirical defenses lack formal robustness guarantees and could be broken by adaptive or unknown attacks. Existing provable defenses have the following limitations: 1) they achieve sub-optimal robustness guarantees for graph structure perturbation, 2) they cannot provide robustness guarantees for arbitrarily node feature perturbations, 3) their robustness guarantees are probabilistic, meaning they could be incorrect with a non-zero probability, and 4) they incur large computation costs. We aim to address those limitations in this work. We propose GNNCert, a certified defense against both graph structure and node feature perturbations for graph classification. Our GNNCert provably predicts the same label for a graph when the number of perturbed edges and the number of nodes with perturbed features are bounded. Our results on 8 benchmark datasets show that GNNCert outperforms three state-of-the-art methods.

Due to the rise of privacy concerns, in many practical applications, the training data is aggregated before being shared with the learner to protect the privacy of users' sensitive responses. In an aggregate learning framework, the dataset is grouped into bags of samples, where each bag is available only with an aggregate response, providing a summary of individuals' responses in that bag. In this paper, we study two natural loss functions for learning from aggregate responses: the bag-level loss and the instance-level loss. In the former, the model is learned by minimizing a loss between the aggregate responses and aggregate model predictions, while in the latter, the model aims to fit individual predictions to the aggregate responses. In this work, we show that the instance-level loss can be perceived as a regularized form of the bag-level loss. This observation allows us to compare the two approaches with respect to the bias and variance of the resulting estimators and to introduce a novel interpolating estimator that combines the two approaches. For linear regression tasks, we provide a precise characterization of the risk of the interpolating estimator in an asymptotic regime where the size of the training set grows in proportion to the feature dimension. Our analysis enables us to theoretically understand the effect of different factors, such as bag size, on the model's prediction risk. Additionally, we propose a mechanism for differentially private learning from aggregate responses and derive the optimal bag size in terms of the prediction risk-privacy trade-off. We also carry out thorough experiments to corroborate our theory and show the efficacy of the interpolating estimator.

Graph Neural Networks (GNNs) have demonstrated state-of-the-art performance in various graph representation learning tasks. Recently, studies revealed their vulnerability to adversarial attacks. In this work, we theoretically define the concept of expected robustness in the context of attributed graphs and relate it to the classical definition of adversarial robustness in the graph representation learning literature. Our definition allows us to derive an upper bound of the expected robustness of Graph Convolutional Networks (GCNs) and Graph Isomorphism Networks subject to node feature attacks. Building on these findings, we connect the expected robustness of GNNs to the orthonormality of their weight matrices and consequently propose an attack-independent, more robust variant of the GCN, called the Graph Convolutional Orthonormal Robust Networks (GCORNs). We further introduce a probabilistic method to estimate the expected robustness, which allows us to evaluate the effectiveness of GCORN on several real-world datasets. Experimental experiments showed that GCORN outperforms available defense methods. Our code is publicly available at: https://github.com/Sennadir/GCORN .

Real-life applications of deep neural networks are hindered by their unsteady predictions when faced with noisy inputs and adversarial attacks. The certified radius in this context is a crucial indicator of the robustness of models. However how to design an efficient classifier with an associated certified radius? Randomized smoothing provides a promising framework by relying on noise injection into the inputs to obtain a smoothed and robust classifier. In this paper, we first show that the variance introduced by the Monte-Carlo sampling in the randomized smoothing procedure estimate closely interacts with two other important properties of the classifier, \textit{i.e.} its Lipschitz constant and margin. More precisely, our work emphasizes the dual impact of the Lipschitz constant of the base classifier, on both the smoothed classifier and the empirical variance. To increase the certified robust radius, we introduce a different way to convert logits to probability vectors for the base classifier to leverage the variance-margin trade-off. We leverage the use of Bernstein's concentration inequality along with enhanced Lipschitz bounds for randomized smoothing. Experimental results show a significant improvement in certified accuracy compared to current state-of-the-art methods. Our novel certification procedure allows us to use pre-trained models with randomized smoothing, effectively improving the current certification radius in a zero-shot manner.

Backdoor attacks pose a significant security risk to machine learning applications due to their stealthy nature and potentially serious consequences. Such attacks involve embedding triggers within a learning model with the intention of causing malicious behavior when an active trigger is present while maintaining regular functionality without it. This paper derives a fundamental understanding of backdoor attacks that applies to both discriminative and generative models, including diffusion models and large language models. We evaluate the effectiveness of any backdoor attack incorporating a constant trigger, by establishing tight lower and upper boundaries for the performance of the compromised model on both clean and backdoor test data. The developed theory answers a series of fundamental but previously underexplored problems, including (1) what are the determining factors for a backdoor attack's success, (2) what is the direction of the most effective backdoor attack, and (3) when will a human-imperceptible trigger succeed. We demonstrate the theory by conducting experiments using benchmark datasets and state-of-the-art backdoor attack scenarios. Our code is available \href{https://github.com/KeyWgh/DemystifyBackdoor}{here}.