Search

Large language models (LLMs) are susceptible to memorizing training data, raising concerns due to the potential extraction of sensitive information. Current methods to measure memorization rates of LLMs, primarily discoverable extraction (Carlini et al., 2022), rely on single-sequence greedy sampling, potentially underestimating the true extent of memorization. This paper introduces a probabilistic relaxation of discoverable extraction that quantifies the probability of extracting a target sequence within a set of generated samples, considering various sampling schemes and multiple attempts. This approach addresses the limitations of reporting memorization rates through discoverable extraction by accounting for the probabilistic nature of LLMs and user interaction patterns. Our experiments demonstrate that this probabilistic measure can reveal cases of higher memorization rates compared to rates found through discoverable extraction. We further investigate the impact of different sampling schemes on extractability, providing a more comprehensive and realistic assessment of LLM memorization and its associated risks. Our contributions include a new probabilistic memorization definition, empirical evidence of its effectiveness, and a thorough evaluation across different models, sizes, sampling schemes, and training data repetitions.

Privacy-preserving machine learning (PPML) enables multiple data owners to contribute their data privately to a set of servers that run a secure multi-party computation (MPC) protocol to train a joint ML model. In these protocols, the input data remains private throughout the training process, and only the resulting model is made available. While this approach benefits privacy, it also exacerbates the risks of data poisoning, where compromised data owners induce undesirable model behavior by contributing malicious datasets. Existing MPC mechanisms can mitigate certain poisoning attacks, but these measures are not exhaustive. To complement existing poisoning defenses, we introduce UTrace: a framework for User-level Traceback of poisoning attacks in PPML. Utrace computes user responsibility scores using gradient similarity metrics aggregated across the most relevant samples in an owner's dataset. UTrace is effective at low poisoning rates and is resilient to poisoning attacks distributed across multiple data owners, unlike existing unlearning-based methods. We introduce methods for checkpointing gradients with low storage overhead, enabling traceback in the absence of data owners at deployment time. We also design several optimizations that reduce traceback time and communication in MPC. We provide a comprehensive evaluation of UTrace across four datasets from three data modalities (vision, text, and malware) and show its effectiveness against 10 poisoning attacks., Comment: 28 pages, 10 figures

Retrieval Augmented Generation (RAG) expands the capabilities of modern large language models (LLMs), by anchoring, adapting, and personalizing their responses to the most relevant knowledge sources. It is particularly useful in chatbot applications, allowing developers to customize LLM output without expensive retraining. Despite their significant utility in various applications, RAG systems present new security risks. In this work, we propose new attack vectors that allow an adversary to inject a single malicious document into a RAG system's knowledge base, and mount a backdoor poisoning attack. We design Phantom, a general two-stage optimization framework against RAG systems, that crafts a malicious poisoned document leading to an integrity violation in the model's output. First, the document is constructed to be retrieved only when a specific trigger sequence of tokens appears in the victim's queries. Second, the document is further optimized with crafted adversarial text that induces various adversarial objectives on the LLM output, including refusal to answer, reputation damage, privacy violations, and harmful behaviors. We demonstrate our attacks on multiple LLM architectures, including Gemma, Vicuna, and Llama, and show that they transfer to GPT-3.5 Turbo and GPT-4. Finally, we successfully conducted a Phantom attack on NVIDIA's black-box production RAG system, "Chat with RTX".