ICLR 2024 Papers

We propose a new approach to promote safety in classification tasks with concept annotations. Our approach – called a *conceptual safeguard* – acts as a verification layer for models that predict a target outcome by first predicting the presence of intermediate concepts. Given this architecture, a safeguard ensures that a model meets a minimal level of accuracy by abstaining from uncertain predictions. In contrast to a standard selective classifier, a safeguard provides an avenue to improve coverage by allowing a human to confirm the presence of uncertain concepts on instances on which it abstains. We develop methods to build safeguards that maximize coverage without compromising safety, namely techniques to propagate the uncertainty in concept predictions and to flag salient concepts for human review. We benchmark our approach on a collection of real-world and synthetic datasets, showing that it can improve performance and coverage in deep learning tasks.

Aligning large language models with human objectives is paramount, yet common approaches including RLHF suffer from unstable and resource-intensive training. In response to this challenge, we introduce ARGS, Alignment as Reward-Guided Search, a novel framework that integrates alignment into the decoding process, eliminating the need for expensive RL training. By adjusting the model’s probabilistic predictions using a reward signal, ARGS generates texts with semantic diversity while being aligned with human preferences, offering a promising and flexible solution for aligning language models. Notably, our method demonstrates consistent enhancements in average reward compared to baselines across diverse alignment tasks and various model dimensions. For example, under the same greedy-based decoding strategy, our method improves the average reward by 19.56% relative to the baseline and secures a preference or tie score of 64.33% in GPT-4 evaluation. We believe that our framework, emphasizing test-time alignment, paves the way for more responsive language models in the future. Code is publicly available at: https://github.com/deeplearning-wisc/args.

Federated learning is an emerging technology for training machine learning models across decentralized data sources without sharing data. Vertical federated learning, also known as feature-based federated learning, applies to scenarios where data sources have the same sample IDs but different feature sets. To ensure fairness among data owners, it is critical to objectively assess the contributions from different data sources and compensate the corresponding data owners accordingly. The Shapley value is a provably fair contribution valuation metric originating from cooperative game theory. However, its straight-forward computation requires extensively retraining a model on each potential combination of data sources, leading to prohibitively high communication and computation overheads due to multiple rounds of federated learning. To tackle this challenge, we propose a contribution valuation metric called vertical federated Shapley value (VerFedSV) based on the classic Shapley value. We show that VerFedSV not only satisfies many desirable properties of fairness but is also efficient to compute. Moreover, VerFedSV can be adapted to both synchronous and asynchronous vertical federated learning algorithms. Both theoretical analysis and extensive experimental results demonstrate the fairness, efficiency, adaptability, and effectiveness of VerFedSV.

Training and deploying machine learning models that meet fairness criteria for protected groups are fundamental in modern artificial intelligence. While numerous constraints and regularization terms have been proposed in the literature to promote fairness in machine learning tasks, most of these approaches are not amenable to stochastic optimization due to the complex and nonlinear structure of constraints and regularizers. Here, the term ``stochastic'' refers to the ability of the algorithm to work with small mini-batches of data. Motivated by the limitation of existing literature, this paper presents a unified stochastic optimization framework for fair empirical risk minimization based on $f$-divergence measures ($f$-FERM). The proposed stochastic algorithm enjoys theoretical convergence guarantees. In addition, our experiments demonstrate the superiority of fairness-accuracy tradeoffs offered by $f$-FERM for almost all batch sizes (ranging from full-batch to batch size of one). Moreover, we show that our framework can be extended to the case where there is a distribution shift from training to the test data. Our extension is based on a distributionally robust optimization reformulation of $f$-FERM objective under $\ell_p$ norms as uncertainty sets. Again, in this distributionally robust setting, $f$-FERM not only enjoys theoretical convergence guarantees but also outperforms other baselines in the literature in the tasks involving distribution shifts. An efficient stochastic implementation of $f$-FERM is publicly available.

Misclassification detection is an important problem in machine learning, as it allows for the identification of instances where the model's predictions are unreliable. However, conventional uncertainty measures such as Shannon entropy do not provide an effective way to infer the real uncertainty associated with the model's predictions. In this paper, we introduce a novel data-driven measure of uncertainty relative to an observer for misclassification detection. By learning patterns in the distribution of soft-predictions, our uncertainty measure can identify misclassified samples based on the predicted class probabilities. Interestingly, according to the proposed measure, soft-predictions corresponding to misclassified instances can carry a large amount of uncertainty, even though they may have low Shannon entropy. We demonstrate empirical improvements over multiple image classification tasks, outperforming state-of-the-art misclassification detection methods.

Group robustness has become a major concern in machine learning (ML) as conventional training paradigms were found to produce high error on minority groups. Without explicit group annotations, proposed solutions rely on heuristics that aim to identify and then amplify the minority samples during training. In our work, we first uncover a critical shortcoming of these methods: an inability to distinguish legitimate minority samples from poison samples in the training set. By amplifying poison samples as well, group robustness methods inadvertently boost the success rate of an adversary---e.g., from 0\% without amplification to over 97\% with it. Notably, we supplement our empirical evidence with an impossibility result proving this inability of a standard heuristic under some assumptions. Moreover, scrutinizing recent poisoning defenses both in centralized and federated learning, we observe that they rely on similar heuristics to identify which samples should be eliminated as poisons. In consequence, minority samples are eliminated along with poisons, which damages group robustness---e.g., from 55\% without the removal of the minority samples to 41\% with it. Finally, as they pursue opposing goals using similar heuristics, our attempt to alleviate the trade-off by combining group robustness methods and poisoning defenses falls short. By exposing this tension, we also hope to highlight how benchmark-driven ML scholarship can obscure the trade-offs among different metrics with potentially detrimental consequences.

The rapid progress in open-source large language models (LLMs) is significantly advancing AI development. Extensive efforts have been made before model release to align their behavior with human values, with the primary goal of ensuring their helpfulness and harmlessness. However, even carefully aligned models can be manipulated maliciously, leading to unintended behaviors, known as ``jailbreaks". These jailbreaks are typically triggered by specific text inputs, often referred to as adversarial prompts. In this work, we propose the generation exploitation attack, an extremely simple approach that disrupts model alignment by only manipulating variations of decoding methods. By exploiting different generation strategies, including varying decoding hyper-parameters and sampling methods, we increase the attack success rate from $0\%$ to more than $95\%$ across 11 language models including LLaMA2, Vicuna, Falcon, and MPT families, outperforming state-of-the-art attacks with $30\times$ lower computational cost. Finally, we propose an effective alignment method that explores diverse generation strategies, which can reasonably reduce the attack success rate under our attack. Altogether, our study underscores a major failure in current safety evaluation and alignment procedures for open-source LLMs, strongly advocating for more comprehensive red teaming and better alignment before releasing such models.

Neural architecture search (NAS) has become a key component of AutoML and a standard tool to automate the design of deep neural networks. Recently, training-free NAS as an emerging paradigm has successfully reduced the search costs of standard training-based NAS by estimating the true architecture performance with only training-free metrics. Nevertheless, the estimation ability of these metrics typically varies across different tasks, making it challenging to achieve robust and consistently good search performance on diverse tasks with only a single training-free metric. Meanwhile, the estimation gap between training-free metrics and the true architecture performances limits training-free NAS to achieve superior performance. To address these challenges, we propose the robustifying and boosting training-free NAS (RoBoT) algorithm which (a) employs the optimized combination of existing training-free metrics explored from Bayesian optimization to develop a robust and consistently better-performing metric on diverse tasks, and (b) applies greedy search, i.e., the exploitation, on the newly developed metric to bridge the aforementioned gap and consequently to boost the search performance of standard training-free NAS further. Remarkably, the expected performance of our RoBoT can be theoretically guaranteed, which improves over the existing training-free NAS under mild conditions with additional interesting insights. Our extensive experiments on various NAS benchmark tasks yield substantial empirical evidence to support our theoretical results.

Learning Rate Rewinding (LRR) has been established as a strong variant of Iterative Magnitude Pruning (IMP) to find lottery tickets in deep overparameterized neural networks. While both iterative pruning schemes couple structure and parameter learning, understanding how LRR excels in both aspects can bring us closer to the design of more flexible deep learning algorithms that can optimize diverse sets of sparse architectures. To this end, we conduct experiments that disentangle the effect of mask learning and parameter optimization and how both benefit from overparameterization. The ability of LRR to flip parameter signs early and stay robust to sign perturbations seems to make it not only more effective in mask identification but also in optimizing diverse sets of masks, including random ones. In support of this hypothesis, we prove in a simplified single hidden neuron setting that LRR succeeds in more cases than IMP, as it can escape initially problematic sign configurations.

The effect of underrepresentation on the performance of minority groups is known to be a serious problem in supervised learning settings; however, it has been underexplored so far in the context of self-supervised learning (SSL). In this paper, we demonstrate that contrastive learning (CL), a popular variant of SSL, tends to collapse representations of minority groups with certain majority groups. We refer to this phenomenon as representation harm and demonstrate it on image and text datasets using the corresponding popular CL methods. Furthermore, our causal mediation analysis of allocation harm on a downstream classification task reveals that representation harm is partly responsible for it, thus emphasizing the importance of studying and mitigating representation harm. Finally, we provide a theoretical explanation for representation harm using a stochastic block model that leads to a representational neural collapse in a contrastive learning setting.

Federated learning has emerged as a promising distributed learning paradigm that facilitates collaborative learning among multiple parties without transferring raw data. However, most existing federated learning studies focus on either horizontal or vertical data settings, where the data of different parties are assumed to be from the same feature or sample space. In practice, a common scenario is the hybrid data setting, where data from different parties may differ both in the features and samples. To address this, we propose HybridTree, a novel federated learning approach that enables federated tree learning on hybrid data. We observe the existence of consistent split rules in trees. With the help of these split rules, we theoretically show that the knowledge of parties can be incorporated into the lower layers of a tree. Based on our theoretical analysis, we propose a layer-level solution that does not need frequent communication traffic to train a tree. Our experiments demonstrate that HybridTree can achieve comparable accuracy to the centralized setting with low computational and communication overhead. HybridTree can achieve up to 8 times speedup compared with the other baselines.

Pre-trained large language models (LLMs) can understand and align with human instructions by supervised fine-tuning (SFT). It is commonly believed that diverse and complex SFT data are of the essence to enable good instruction-following abilities. However, such diversity and complexity are obscure and lack quantitative analyses. In this work, we propose InsTag, an open-set instruction tagging method, to identify semantics and intentions of human instructions by tags that provide access to definitions and quantified analyses of instruction diversity and complexity. We obtain 6.6K fine-grained tags to describe instructions from popular open-sourced SFT datasets comprehensively. We find that the abilities of aligned LLMs benefit from more diverse and complex instructions in SFT data. Based on this observation, we propose a data sampling procedure based on InsTag, and select 6K diverse and complex samples from open-source datasets for SFT. The resulting models, TagLM, outperform open-source models based on considerably larger SFT data evaluated by MT-Bench, echoing the importance of instruction diversity and complexity and the effectiveness of InsTag. InsTag has robust potential to be extended to more applications beyond the data selection as it provides an effective way to analyze the distribution of instructions.

We introduce new jailbreak attacks on vision language models (VLMs), which use aligned LLMs and are resilient to text-only jailbreak attacks. Specifically, we develop cross-modality attacks on alignment where we pair adversarial images going through the vision encoder with textual prompts to break the alignment of the language model. Our attacks employ a novel compositional strategy that combines an image, adversarially targeted towards toxic embeddings, with generic prompts to accomplish the jailbreak. Thus, the LLM draws the context to answer the generic prompt from the adversarial image. The generation of benign-appearing adversarial images leverages a novel embedding-space-based methodology, operating with no access to the LLM model. Instead, the attacks require access only to the vision encoder and utilize one of our four embedding space targeting strategies. By not requiring access to the LLM, the attacks lower the entry barrier for attackers, particularly when vision encoders such as CLIP are embedded in closed-source LLMs. The attacks achieve a high success rate across different VLMs, highlighting the risk of cross-modality alignment vulnerabilities, and the need for new alignment approaches for multi-modal models.

As of September 2023, ChatGPT correctly answers “what is 7+8” with 15, but when asked “7+8=15, True or False” it responds with “False”. This inconsistency between generating and validating an answer is prevalent in language models (LMs) and erodes trust. In this paper, we propose a framework for measuring the consistency between generation and validation (which we call generator-validator consistency, or GV-consistency), finding that even GPT-4 (0613), a state-of-the-art LM, is GV-consistent only 76% of the time. To improve the consistency of LMs, we propose to finetune on the filtered generator and validator responses that are GV-consistent, and call this approach consistency fine-tuning. We find that this approach improves GV-consistency of Alpaca-30B from 60% to 93%, and the improvement extrapolates to unseen tasks and domains (e.g., GV-consistency for positive style transfers extrapolates to unseen styles like humor). In addition to improving consistency, consistency fine-tuning improves both generator quality and validator accuracy without using any labeled data. Evaluated across 6 tasks, including math questions, knowledge-intensive QA, and instruction following, our method improves generator quality by an average of 16% and validator accuracy by an average of 6.3% across all tasks.

Large language models (LLMs) often demonstrate inconsistencies with human preferences. Previous research typically gathered human preference data and then aligned the pre-trained models using reinforcement learning or instruction tuning, a.k.a. the finetuning step. In contrast, aligning frozen LLMs without requiring alignment data is more appealing. This work explores the potential of the latter setting. We discover that by integrating self-evaluation and rewind mechanisms, unaligned LLMs can directly produce responses consistent with human preferences via self-boosting. We introduce a novel inference method, Rewindable Auto-regressive INference (RAIN), that allows pre-trained LLMs to evaluate their own generation and use the evaluation results to guide rewind and generation for AI safety. Notably, RAIN operates without the need of extra data for model alignment and abstains from any training, gradient computation, or parameter updates. Experimental results evaluated by GPT-4 and humans demonstrate the effectiveness of RAIN: on the HH dataset, RAIN improves the harmlessness rate of LLaMA 30B from 82% of vanilla inference to 97%, while maintaining the helpfulness rate. On the TruthfulQA dataset, RAIN improves the truthfulness of the already-well-aligned LLaMA-2-chat 13B model by 5%.

Contrastive Learning (CL) has attracted enormous attention due to its remarkable capability in unsupervised representation learning. However, recent works have revealed the vulnerability of CL to backdoor attacks: the feature extractor could be misled to embed backdoored data close to an attack target class, thus fooling the downstream predictor to misclassify it as the target. Existing attacks usually adopt a fixed trigger pattern and poison the training set with trigger-injected data, hoping for the feature extractor to learn the association between trigger and target class. However, we find that such fixed trigger design fails to effectively associate trigger-injected data with target class in the embedding space due to special CL mechanisms, leading to a limited attack success rate (ASR). This phenomenon motivates us to find a better backdoor trigger design tailored for CL framework. In this paper, we propose a bi-level optimization approach to achieve this goal, where the inner optimization simulates the CL dynamics of a surrogate victim, and the outer optimization enforces the backdoor trigger to stay close to the target throughout the surrogate CL procedure. Extensive experiments show that our attack can achieve a higher attack success rate (e.g., 99\% ASR on ImageNet-100) with a very low poisoning rate (1\%). Besides, our attack can effectively evade existing state-of-the-art defenses.

Adversarial patch attacks, where a small patch is placed in the scene to fool neural networks, have been studied for numerous applications. Focusing on image classification, we consider the setting of a black-box transfer attack where an attacker does not know the target model. Instead of forcing corrupted image representations to cross the nearest decision boundaries or converge to a particular point, we propose a distribution-oriented approach. We rely on optimal transport to push the feature distribution of attacked images towards an already modeled distribution. We show that this new distribution-oriented approach leads to better transferable patches. Through digital experiments conducted on ImageNet-1K, we provide evidence that our new patches are the only ones that can simultaneously influence multiple Transformer models and Convolutional Neural Networks. Physical world experiments demonstrate that our patch can affect systems in deployment without explicit knowledge.

Following the success of Large Language Models (LLMs), Large Multimodal Models (LMMs), such as the Flamingo model and its subsequent competitors, have started to emerge as natural steps towards generalist agents. However, interacting with recent LMMs reveals major limitations that are hardly captured by the current evaluation benchmarks. Indeed, task performances (e.g., VQA accuracy) alone do not provide enough clues to understand their real capabilities, limitations, and to which extent such models are aligned to human expectations. To refine our understanding of those flaws, we deviate from the current evaluation paradigm, and (1) evaluate 10 recent open-source LMMs from 3B up to 80B parameter scale, on 5 different axes; hallucinations, abstention, compositionality, explainability and instruction following. Our evaluation on these axes reveals major flaws in LMMs. While the current go-to solution to align these models is based on training, such as instruction tuning or RLHF, we rather (2) explore the training-free in-context learning (ICL) as a solution, and study how it affects these limitations. Based on our ICL study, (3) we push ICL further and propose new multimodal ICL variants such as; Multitask-ICL, Chain-of-Hindsight-ICL, and Self-Correcting-ICL. Our findings are as follows; (1) Despite their success, LMMs have flaws that remain unsolved with scaling alone. (2) The effect of ICL on LMMs flaws is nuanced; despite its effectiveness for improved explainability, answer abstention, ICL only slightly improves instruction following, does not improve compositional abilities, and actually even amplifies hallucinations. (3) The proposed ICL variants are promising as post-hoc approaches to efficiently tackle some of those flaws. The code is available here: https://github.com/mshukor/EvALign-ICL.

Direct image alignment is a widely used technique for relative 6DoF pose estimation between two images, but its accuracy strongly depends on pose initialization. Therefore, recent end-to-end frameworks increase the convergence basin of the learned feature descriptors with special training objectives, such as the Gauss-Newton loss. However, the training data may exhibit bias toward a specific type of motion and pose initialization, thus limiting the generalization of these methods. In this work, we derive a closed-form solution to the expected optimum of the Gauss-Newton loss. The solution is agnostic to the underlying feature representation and allows us to dynamically adjust the basin of convergence according to our assumptions about the uncertainty in the current estimates. These properties allow for effective control over the convergence in the alignment process. Despite using self-supervised feature embeddings, our solution achieves compelling accuracy w.r.t. the state-of-the-art direct image alignment methods trained end-to-end with pose supervision, and demonstrates improved robustness to pose initialization. Our analytical solution exposes some inherent limitations of end-to-end learning with the Gauss-Newton loss, and establishes an intriguing connection between direct image alignment and feature-matching approaches.