Reinforcement Learning from Human Feedback (RLHF) is a popular method for aligning Language Models (LM) with human values and preferences. RLHF requires a large number of preference pairs as training data, which are often used in both the Supervised Fine-Tuning and Reward Model training and therefore publicly available datasets are commonly used. In this work, we study to what extent a malicious actor can manipulate the LMs generations by poisoning the preferences, i.e., injecting poisonous preference pairs into these datasets and the RLHF training process. We propose strategies to build poisonous preference pairs and test their performance by poisoning two widely used preference datasets. Our results show that preference poisoning is highly effective: injecting a small amount of poisonous data (1-5\% of the original dataset), we can effectively manipulate the LM to generate a target entity in a target sentiment (positive or negative). The findings from our experiments also shed light on strategies to defend against the preference poisoning attack.
相關內容
Decision Tree (DT) Learning is a fundamental problem in Interpretable Machine Learning, yet it poses a formidable optimisation challenge. Despite numerous efforts dating back to the early 1990's, practical algorithms have only recently emerged, primarily leveraging Dynamic Programming (DP) and Branch & Bound (B&B) techniques. These methods fall into two categories: algorithms like DL8.5, MurTree and STreeD utilise an efficient DP strategy but lack effective bounds for pruning the search space; while algorithms like OSDT and GOSDT employ more efficient pruning bounds but at the expense of a less refined DP strategy. We introduce Branches, a new algorithm that combines the strengths of both approaches. Using DP and B&B with a novel analytical bound for efficient pruning, Branches offers both speed and sparsity optimisation. Unlike other methods, it also handles non-binary features. Theoretical analysis shows its lower complexity compared to existing methods, and empirical results confirm that Branches outperforms the state-of-the-art in speed, iterations, and optimality.
RowHammer is a major read disturbance mechanism in DRAM where repeatedly accessing (hammering) a row of DRAM cells (DRAM row) induces bitflips in other physically nearby DRAM rows. RowHammer solutions perform preventive actions (e.g., refresh neighbor rows of the hammered row) that mitigate such bitflips to preserve memory isolation, a fundamental building block of security and privacy in modern computing systems. However, preventive actions induce non-negligible memory request latency and system performance overheads as they interfere with memory requests. As shrinking technology node size over DRAM chip generations exacerbates RowHammer, the overheads of RowHammer solutions become prohibitively expensive. As a result, a malicious program can effectively hog the memory system and deny service to benign applications by causing many RowHammer-preventive actions. In this work, we tackle the performance overheads of RowHammer solutions by tracking and throttling the generators of memory accesses that trigger RowHammer solutions. To this end, we propose BreakHammer. BreakHammer 1) observes the time-consuming RowHammer-preventive actions of existing RowHammer mitigation mechanisms, 2) identifies hardware threads that trigger many of these actions, and 3) reduces the memory bandwidth usage of each identified thread. As such, BreakHammer significantly reduces the number of RowHammer-preventive actions performed, thereby improving 1) system performance and DRAM energy, and 2) reducing the maximum slowdown induced on a benign application, with near-zero area overhead. Our extensive evaluations demonstrate that BreakHammer effectively reduces the negative performance, energy, and fairness effects of eight RowHammer mitigation mechanisms. To foster further research we open-source our BreakHammer implementation and scripts at //github.com/CMU-SAFARI/BreakHammer.
Fill-in-the-Middle (FIM) has become integral to code language models, enabling generation of missing code given both left and right contexts. However, the current FIM training paradigm, which reorders original training sequences and then performs regular next-token prediction (NTP), often leads to models struggling to generate content that aligns smoothly with the surrounding context. Crucially, while existing works rely on rule-based post-processing to circumvent this weakness, such methods are not practically usable in open-domain code completion tasks as they depend on restrictive, dataset-specific assumptions (e.g., generating the same number of lines as in the ground truth). Moreover, model performance on FIM tasks deteriorates significantly without these unrealistic assumptions. We hypothesize that NTP alone is insufficient for models to learn effective planning conditioned on the distant right context, a critical factor for successful code infilling. To overcome this, we propose Horizon-Length Prediction (HLP), a novel training objective that teaches models to predict the number of remaining middle tokens (i.e., horizon length) at each step. HLP advances FIM with lookahead planning, enabling models to inherently learn infilling boundaries for arbitrary left and right contexts without relying on dataset-specific post-processing. Our evaluation across different models and sizes shows that HLP significantly improves FIM performance by up to 24% relatively on diverse benchmarks, across file-level and repository-level, and without resorting to unrealistic post-processing methods. Furthermore, the enhanced planning capability gained through HLP boosts model performance on code reasoning. Importantly, HLP only incurs negligible training overhead and no additional inference cost, ensuring its practicality for real-world scenarios.
Classification tasks are typically handled using Machine Learning (ML) models, which lack a balance between accuracy and interpretability. This paper introduces a new approach to using Large Language Models (LLMs) for classification tasks in an explainable way. Unlike ML models that rely heavily on data cleaning and feature engineering, this method streamlines the process using LLMs. This paper proposes a new concept called "Language Model Learning (LML)" powered by a new method called "Data-Augmented Prediction (DAP)". The classification is performed by LLMs using a method similar to humans manually exploring and understanding the data and deciding classifications using data as a reference. In the LML process, a dataset is summarized and evaluated to determine the features that lead to the classification of each label the most. In the process of DAP, the system uses the data summary and a row of the testing dataset to automatically generate a query, which is used to retrieve relevant rows from the dataset. A classification is generated by the LLM using data summary and relevant rows, ensuring satisfactory accuracy even with complex data using context-aware decision-making. LML and DAP unlock the possibilities of new applications. The proposed method uses the words "Act as an Explainable Machine Learning Model" in the prompt to enhance the interpretability of the predictions by allowing users to review the logic behind each prediction. In some test cases, the system scored an accuracy above 90%, proving the effectiveness of the system and its potential to outperform conventional ML models in various scenarios. The code is available at //github.com/Pro-GenAI/LML-DAP
Large language models (LLMs), while exhibiting exceptional performance, suffer from hallucinations, especially on knowledge-intensive tasks. Existing works propose to augment LLMs with individual text units retrieved from external knowledge corpora to alleviate the issue. However, in many domains, texts are interconnected (e.g., academic papers in a bibliographic graph are linked by citations and co-authorships) which form a (text-attributed) graph. The knowledge in such graphs is encoded not only in single texts/nodes but also in their associated connections. To facilitate the research of augmenting LLMs with graphs, we manually construct a Graph Reasoning Benchmark dataset called GRBench, containing 1,740 questions that can be answered with the knowledge from 10 domain graphs. Then, we propose a simple and effective framework called Graph Chain-of-thought (Graph-CoT) to augment LLMs with graphs by encouraging LLMs to reason on the graph iteratively. Each Graph-CoT iteration consists of three sub-steps: LLM reasoning, LLM-graph interaction, and graph execution. We conduct systematic experiments with three LLM backbones on GRBench, where Graph-CoT outperforms the baselines consistently. The code is available at //github.com/PeterGriffinJin/Graph-CoT.
Current language models use subword-based tokenization algorithms like Byte Pair Encoding, which put their validity as models of linguistic representations into question. In this paper, we explore the potential of tokenization-free, phoneme- and grapheme-based language models. We demonstrate that small models based on the Llama architecture can achieve strong linguistic performance on standard syntactic and novel lexical/phonetic benchmarks when trained with character-level vocabularies. We further show that phoneme-based models without any graphemic biases almost match grapheme-based models in standard tasks and novel evaluations. Our findings suggest a promising direction for creating more linguistically plausible language models that are better suited for computational studies of language acquisition and processing.
Causal Machine Learning (CausalML) is an umbrella term for machine learning methods that formalize the data-generation process as a structural causal model (SCM). This allows one to reason about the effects of changes to this process (i.e., interventions) and what would have happened in hindsight (i.e., counterfactuals). We categorize work in \causalml into five groups according to the problems they tackle: (1) causal supervised learning, (2) causal generative modeling, (3) causal explanations, (4) causal fairness, (5) causal reinforcement learning. For each category, we systematically compare its methods and point out open problems. Further, we review modality-specific applications in computer vision, natural language processing, and graph representation learning. Finally, we provide an overview of causal benchmarks and a critical discussion of the state of this nascent field, including recommendations for future work.
Deep Learning (DL) is vulnerable to out-of-distribution and adversarial examples resulting in incorrect outputs. To make DL more robust, several posthoc anomaly detection techniques to detect (and discard) these anomalous samples have been proposed in the recent past. This survey tries to provide a structured and comprehensive overview of the research on anomaly detection for DL based applications. We provide a taxonomy for existing techniques based on their underlying assumptions and adopted approaches. We discuss various techniques in each of the categories and provide the relative strengths and weaknesses of the approaches. Our goal in this survey is to provide an easier yet better understanding of the techniques belonging to different categories in which research has been done on this topic. Finally, we highlight the unsolved research challenges while applying anomaly detection techniques in DL systems and present some high-impact future research directions.
The problem of Multiple Object Tracking (MOT) consists in following the trajectory of different objects in a sequence, usually a video. In recent years, with the rise of Deep Learning, the algorithms that provide a solution to this problem have benefited from the representational power of deep models. This paper provides a comprehensive survey on works that employ Deep Learning models to solve the task of MOT on single-camera videos. Four main steps in MOT algorithms are identified, and an in-depth review of how Deep Learning was employed in each one of these stages is presented. A complete experimental comparison of the presented works on the three MOTChallenge datasets is also provided, identifying a number of similarities among the top-performing methods and presenting some possible future research directions.
State-of-the-art Convolutional Neural Network (CNN) benefits a lot from multi-task learning (MTL), which learns multiple related tasks simultaneously to obtain shared or mutually related representations for different tasks. The most widely-used MTL CNN structure is based on an empirical or heuristic split on a specific layer (e.g., the last convolutional layer) to minimize different task-specific losses. However, this heuristic sharing/splitting strategy may be harmful to the final performance of one or multiple tasks. In this paper, we propose a novel CNN structure for MTL, which enables automatic feature fusing at every layer. Specifically, we first concatenate features from different tasks according to their channel dimension, and then formulate the feature fusing problem as discriminative dimensionality reduction. We show that this discriminative dimensionality reduction can be done by 1x1 Convolution, Batch Normalization, and Weight Decay in one CNN, which we refer to as Neural Discriminative Dimensionality Reduction (NDDR). We perform ablation analysis in details for different configurations in training the network. The experiments carried out on different network structures and different task sets demonstrate the promising performance and desirable generalizability of our proposed method.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191