Searcharxiv⌕ Search

arXiv · 2611.15013

Protecting Sensitive Knowledge Under LLM Unlearning: The Proxy Threat and a Superposition-Aware Defense

Abstract

Machine unlearning lets the owner of large language models (LLMs) remove specific entities from a trained model without retraining from scratch. Existing machine unlearning methods focus on removing the requested entities while maintaining utility. However, we observe that the representations of distinct entities in an LLM may overlap in shared parameter subspaces, a phenomenon known as knowledge superposition. Removing one entity may then affect other entities. This exposes a new security threat: legitimate unlearning requests can be exploited to remove sensitive entities that were never requested. This threat is unlike data poisoning or backdoor attacks. We define the attack as the Proxy Unlearning Attack (PUA), which selects removable proxy entities superposed with a sensitive target. The attack can even occur unintentionally, when benign users accidentally remove entities superposed with a sensitive one. In response, we propose Superposition-Aware Unlearning Protection (SAUP), a general defense framework that can be applied to unlearning requests to ensure a correct removal. SAUP performs in two stages. Before unlearning, it uses superposition detection to estimate which sensitive entities the request puts at risk. After unlearning, it decouples the shared representations and restores the affected entities, while keeping the requested entities suppressed. Our findings reveal that legitimate unlearning requests can be maliciously exploited through knowledge superposition. When dealing with unlearning requests, LLM providers need representation-level detection and recovery mechanisms to mitigate this threat and protect sensitive knowledge.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jiayang Li, Tianqing Zhu, Longxiang Gao, Huajie Chen, Dayong Ye, Shang Wang, Huan Huo, Wanlei Zhou. 2026-11-11. Protecting Sensitive Knowledge Under LLM Unlearning: The Proxy Threat and a Superposition-Aware Defense. https://searcharxiv.com/abs/2611.15013

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

HE-LRM: Encrypted Deep Learning Recommendation Models using Fully Homomorphic Encryption

FHE inference systems primarily target models with dense inputs, while Deep Learning Recommendation Models (DLRMs) contain sparse categorical features that require lookups into embedding tables. Encrypted lookups are challenging because indexing is not natively supported and must be expressed using SIMD arithmetic. A direct one-hot formulation does not scale. Prior work addresses parts of this problem through embedding compression or compact encrypted queries. However, it does not address complete DLRM inference, leaving open how private lookup interacts with downstream computation and memory constraints. We present HE-LRM, the first end-to-end FHE system for private DLRM inference. We use deterministic compositional embedding compression with client-side digit decomposition, avoiding expensive homomorphic selector construction while consuming only one multiplicative level. We further develop a block-diagonal packing strategy that performs lookups across multiple embedding tables using optimized BSGS linear transformations and produces encrypted embeddings directly compatible with downstream inference. We extend Orion with a high-performance GPU backend and achieve 9 ms for single-inference embedding lookup. We also show that GPU memory capacity limits batching and therefore the latency-bound throughput in such systems. On Criteo, HE-LRM achieves complete encrypted inference with server-side latencies of 154-247 ms on a single GPU. We open-source HE-LRM at https://github.com/baahl-nyu/orion

cs.CR↗

PSM: Prompt Sensitivity Minimization via LLM-Guided Black-Box Optimization

System prompts are critical for guiding the behavior of Large Language Models (LLMs), yet they often contain proprietary logic or sensitive information, making them a prime target for extraction attacks. Adversarial queries can successfully elicit these hidden instructions, posing significant security and privacy risks. Existing defense mechanisms frequently rely on heuristics, incur substantial computational overhead, or are inapplicable to models accessed via black-box APIs. This paper introduces a novel framework for hardening system prompts through shield appending, a lightweight approach that adds a protective textual layer to the original prompt. Our core contribution is the formalization of prompt hardening as a utility-constrained optimization problem. We leverage an LLM-as-optimizer to search the space of possible SHIELDs, seeking to minimize a leakage metric derived from a suite of adversarial attacks, while simultaneously preserving task utility above a specified threshold, measured by semantic fidelity to baseline outputs. This black-box, optimization-driven methodology is lightweight and practical, requiring only API access to the target and optimizer LLMs. We demonstrate empirically that our optimized SHIELDs significantly reduce prompt leakage against a comprehensive set of extraction attacks, outperforming established baseline defenses without compromising the model's intended functionality. Our work presents a paradigm for developing robust, utility-aware defenses in the escalating landscape of LLM security. The code is made public on the following link: https://github.com/psm-defense/psm

cs.CR↗

Arithmetic limiting laws over shifted primes: Hausdorff dimension zero, Fourier decay, and absolute continuity

We study the classical limiting laws of $φ(n)/n$ and $σ(n)/n$ over integers and shifted primes $p-1$. Let $μ_G$ denote the shifted-prime totient law and $μ_f$ its image under $x\mapsto-\log x$. The law $μ_G$ has Hausdorff dimension zero, while every Borel set of positive $μ_G$-measure has packing dimension one. All four ratio laws and their logarithmic images have lower and upper Shannon information dimensions satisfying $0.550<\underline{d}_1\le\overline{d}_1<0.571$. The law $μ_G$ gives full mass to absolutely normal Liouville numbers. Both $μ_G$ and $μ_f$ have quantitative Fourier decay. For $μ_f$, we obtain sharp concentration asymptotics and localize maximizing intervals. For $U_T$ uniform on $[0,T]$, $(\log T)\widehat{μ_f}(U_T)$ converges in distribution to a random Euler product with an independent uniform phase. For a uniform square matrix over $\mathbb{F}_p$, the probability of a primitive-root determinant tends, as the matrix size grows, to $c(p)=\frac{φ(p-1)}{p-1}\prod_{j\ge1}(1-p^{-j})$. Over primes, $c(p)$ has limiting law $μ_G$. In every fixed reduced progression, the empirical laws of $c(p)$ and $φ(p-1)/(p-1)$ over primes $p\le x$ converge to their common progression law with sharp Kolmogorov order $1/\log x$. We also obtain transport estimates in growing progressions. Strongly additive functions with prime values $(\log\ell)^{-A}$ have absolutely continuous integer and shifted-prime laws for every $A>0$, disproving the Erdős--Kátai singularity conjecture for $A>2$. Applications include extremal determinant bounds and certified prime selection without complete factorization of $p-1$. These results are unconditional; further quantitative refinements assume the generalized Riemann hypothesis for Dirichlet $L$-functions.

cs.CR↗