Abstract:Concept-based interpretability methods offer a lens into the internals of foundation models by decomposing their embeddings into high-level concepts. These concept representations are most useful when they are compositional, meaning that the individual concepts compose to explain the full sample. We show that existing unsupervised concept extraction methods find concepts which are not compositional. To automatically discover compositional concept representations, we identify two salient properties of such representations, and propose Compositional Concept Extraction (CCE) for finding concepts which obey these properties. We evaluate CCE on five different datasets over image and text data. Our evaluation shows that CCE finds more compositional concept representations than baselines and yields better accuracy on four downstream classification tasks. Code and data are available at https://github.com/adaminsky/compositional_concepts .
Abstract:We study how to subvert language models from following the rules. We model rule-following as inference in propositional Horn logic, a mathematical system in which rules have the form "if $P$ and $Q$, then $R$" for some propositions $P$, $Q$, and $R$. We prove that although transformers can faithfully abide by such rules, maliciously crafted prompts can nevertheless mislead even theoretically constructed models. Empirically, we find that attacks on our theoretical models mirror popular attacks on large language models. Our work suggests that studying smaller theoretical models can help understand the behavior of large language models in rule-based settings like logical reasoning and jailbreak attacks.
Abstract:Pre-trained Large Language Models (LLMs) have demonstrated remarkable capabilities but also pose risks by learning and generating copyrighted material, leading to significant legal and ethical concerns. To address these issues, it is critical for model owners to be able to unlearn copyrighted content at various time steps. We explore the setting of sequential unlearning, where copyrighted content is removed over multiple time steps - a scenario that has not been rigorously addressed. To tackle this challenge, we propose Stable Sequential Unlearning (SSU), a novel unlearning framework for LLMs, designed to have a more stable process to remove copyrighted content from LLMs throughout different time steps using task vectors, by incorporating additional random labeling loss and applying gradient-based weight saliency mapping. Experiments demonstrate that SSU finds a good balance between unlearning efficacy and maintaining the model's general knowledge compared to existing baselines.
Abstract:Many computational tasks can be naturally expressed as a composition of a DNN followed by a program written in a traditional programming language or an API call to an LLM. We call such composites "neural programs" and focus on the problem of learning the DNN parameters when the training data consist of end-to-end input-output labels for the composite. When the program is written in a differentiable logic programming language, techniques from neurosymbolic learning are applicable, but in general, the learning for neural programs requires estimating the gradients of black-box components. We present an algorithm for learning neural programs, called ISED, that only relies on input-output samples of black-box components. For evaluation, we introduce new benchmarks that involve calls to modern LLMs such as GPT-4 and also consider benchmarks from the neurosymolic learning literature. Our evaluation shows that for the latter benchmarks, ISED has comparable performance to state-of-the-art neurosymbolic frameworks. For the former, we use adaptations of prior work on gradient approximations of black-box components as a baseline, and show that ISED achieves comparable accuracy but in a more data- and sample-efficient manner.
Abstract:Designing faithful yet accurate AI models is challenging, particularly in the field of individual treatment effect estimation (ITE). ITE prediction models deployed in critical settings such as healthcare should ideally be (i) accurate, and (ii) provide faithful explanations. However, current solutions are inadequate: state-of-the-art black-box models do not supply explanations, post-hoc explainers for black-box models lack faithfulness guarantees, and self-interpretable models greatly compromise accuracy. To address these issues, we propose DISCRET, a self-interpretable ITE framework that synthesizes faithful, rule-based explanations for each sample. A key insight behind DISCRET is that explanations can serve dually as database queries to identify similar subgroups of samples. We provide a novel RL algorithm to efficiently synthesize these explanations from a large search space. We evaluate DISCRET on diverse tasks involving tabular, image, and text data. DISCRET outperforms the best self-interpretable models and has accuracy comparable to the best black-box models while providing faithful explanations. DISCRET is available at https://github.com/wuyinjun-1993/DISCRET-ICML2024.
Abstract:Jailbreak attacks cause large language models (LLMs) to generate harmful, unethical, or otherwise objectionable content. Evaluating these attacks presents a number of challenges, which the current collection of benchmarks and evaluation techniques do not adequately address. First, there is no clear standard of practice regarding jailbreaking evaluation. Second, existing works compute costs and success rates in incomparable ways. And third, numerous works are not reproducible, as they withhold adversarial prompts, involve closed-source code, or rely on evolving proprietary APIs. To address these challenges, we introduce JailbreakBench, an open-sourced benchmark with the following components: (1) a new jailbreaking dataset containing 100 unique behaviors, which we call JBB-Behaviors; (2) an evolving repository of state-of-the-art adversarial prompts, which we refer to as jailbreak artifacts; (3) a standardized evaluation framework that includes a clearly defined threat model, system prompts, chat templates, and scoring functions; and (4) a leaderboard that tracks the performance of attacks and defenses for various LLMs. We have carefully considered the potential ethical implications of releasing this benchmark, and believe that it will be a net positive for the community. Over time, we will expand and adapt the benchmark to reflect technical and methodological advances in the research community.
Abstract:Aligned large language models (LLMs) are vulnerable to jailbreaking attacks, which bypass the safeguards of targeted LLMs and fool them into generating objectionable content. While initial defenses show promise against token-based threat models, there do not exist defenses that provide robustness against semantic attacks and avoid unfavorable trade-offs between robustness and nominal performance. To meet this need, we propose SEMANTICSMOOTH, a smoothing-based defense that aggregates the predictions of multiple semantically transformed copies of a given input prompt. Experimental results demonstrate that SEMANTICSMOOTH achieves state-of-the-art robustness against GCG, PAIR, and AutoDAN attacks while maintaining strong nominal performance on instruction following benchmarks such as InstructionFollowing and AlpacaEval. The codes will be publicly available at https://github.com/UCSB-NLP-Chang/SemanticSmooth.
Abstract:With the prevalence of the Pretraining-Finetuning paradigm in transfer learning, the robustness of downstream tasks has become a critical concern. In this work, we delve into adversarial robustness in transfer learning and reveal the critical role of initialization, including both the pretrained model and the linear head. First, we discover the necessity of an adversarially robust pretrained model. Specifically, we reveal that with a standard pretrained model, Parameter-Efficient Finetuning~(PEFT) methods either fail to be adversarially robust or continue to exhibit significantly degraded adversarial robustness on downstream tasks, even with adversarial training during finetuning. Leveraging a robust pretrained model, surprisingly, we observe that a simple linear probing can outperform full finetuning and other PEFT methods with random initialization on certain datasets. We further identify that linear probing excels in preserving robustness from the robust pretraining. Based on this, we propose Robust Linear Initialization~(RoLI) for adversarial finetuning, which initializes the linear head with the weights obtained by adversarial linear probing to maximally inherit the robustness from pretraining. Across five different image classification datasets, we demonstrate the effectiveness of RoLI and achieve new state-of-the-art results.
Abstract:An explanation of a machine learning model is considered "faithful" if it accurately reflects the model's decision-making process. However, explanations such as feature attributions for deep learning are not guaranteed to be faithful, and can produce potentially misleading interpretations. In this work, we develop Sum-of-Parts (SOP), a class of models whose predictions come with grouped feature attributions that are faithful-by-construction. This model decomposes a prediction into an interpretable sum of scores, each of which is directly attributable to a sparse group of features. We evaluate SOP on benchmarks with standard interpretability metrics, and in a case study, we use the faithful explanations from SOP to help astrophysicists discover new knowledge about galaxy formation.
Abstract:With evolving data regulations, machine unlearning (MU) has become an important tool for fostering trust and safety in today's AI models. However, existing MU methods focusing on data and/or weight perspectives often grapple with limitations in unlearning accuracy, stability, and cross-domain applicability. To address these challenges, we introduce the concept of 'weight saliency' in MU, drawing parallels with input saliency in model explanation. This innovation directs MU's attention toward specific model weights rather than the entire model, improving effectiveness and efficiency. The resultant method that we call saliency unlearning (SalUn) narrows the performance gap with 'exact' unlearning (model retraining from scratch after removing the forgetting dataset). To the best of our knowledge, SalUn is the first principled MU approach adaptable enough to effectively erase the influence of forgetting data, classes, or concepts in both image classification and generation. For example, SalUn yields a stability advantage in high-variance random data forgetting, e.g., with a 0.2% gap compared to exact unlearning on the CIFAR-10 dataset. Moreover, in preventing conditional diffusion models from generating harmful images, SalUn achieves nearly 100% unlearning accuracy, outperforming current state-of-the-art baselines like Erased Stable Diffusion and Forget-Me-Not.