acceptodds
Under review as a conference paper at ICLR 2027

A Graph-theoretic Approach to Privacy and Robustness in LLMs

Abstract

Large language models (LLMs) face security risks from both unintended disclosure and adversarial manipulation: they may reproduce sensitive training content, leak membership information, and exhibit substantial performance degradation under small targeted prompt perturbations. Although these security threats differ in their manifestation, they are influenced by how an LLM internally routes attention across tokens. Existing approaches lack a common, training-free framework to study LLM security by analyzing and intervening on this internal process. We adopt a common graph-theoretic lens, representing the causally masked attention map of a decoder-only LLM as a directed acyclic graph, with tokens as nodes and attention scores as weighted edges. Using this representation, we introduce two complementary inference-time approaches for analyzing LLM security vulnerabilities: an edge-centric intervention on attention weights for privacy and a node-centric intervention for context robustness. First, AttentionSwap identifies structural agreements in dominant edges of a primary model and an auxiliary model before swapping their weights to disrupt verbatim memorization while preserving utility on the downstream tasks. Second, BotDrop identifies topologically critical tokens and prunes them to stress-test prompt sensitivity. Extensive empirical evaluations demonstrate the effectiveness of this shared graph-based perspective for addressing different LLM security threats through our novel training-free, inference-time interventions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLMs, Attention, Robustness, Memorization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_f9s1OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors proposed two inference-time methods for LLM privacy and security. For privacy, they considered MIA risk; for security, they considered robustness. Both issues are formulated as graph problems. The authors also conduct experiments and provide ablation studies.

Critical strengths

None.

Critical weaknesses

The evaluation is weak and lacks comparisons with SOTA methods (except TokenSwap).

Questions for authors

Why were only simple baselines considered, and why are there so few attack methods?

Reviewer_a5nEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on LLM privacy risks. Specifically, the paper adopts a common graph-theoretic perspective and proposes two complementary inference-time methods. The first method, ATTENTIONSWAP, identifies structural agreements, and the second method, BOTDROP, identifies topologically critical tokens. Several experiments have been conducted to show their efficacy.

Critical strengths
  1. The idea of leveraging graph theory for LLM memorization is interesting.
  2. The paper is clearly presented and easy to follow.
Critical weaknesses
  1. No undefended AUROC is reported; MIMIR is already near chance, and ATTS-AUX rarely beats the random baselines on privacy.
  2. Utility relies on a hidden-state cosine proxy, and the text overlooks a large SciQ drop (0.88 → 0.57) in Table 1, next to a chance-level TokenSwap baseline.
  3. The BotDrop evaluation tests only k = 1 on small sets, omits standard baselines, and may just delete essential content such as numbers in GSM8K.
  4. The two methods share only the attention-as-DAG view, with no theory and arbitrary design choices.
Questions for authors

Please refer to the Critical Weaknesses section.

Reviewer_qEhfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper formulates LLM security analysis through a common graph-theoretic abstraction of internal information routing. Based on this formulation, the authors consider privacy mitigation as an edge-centric intervention problem, and robustness analysis as a node-centric intervention problem. They proposed two training-free, inference-time methods to 1) reduce memorization and MIA leakage, and 2) degrade target model's utility.

Critical strengths

Moving the TokenSwap intuition from output logits into attention routing is interesting. The experiments on AttentionSwap show that it can reduce memorization while maintaining accuracy essentially unchanged, and AttentionSwap can be combined with TokenSwap.

Critical weaknesses
  1. The MIA results do not include undefended AUROC. Tables 2, 6, and 7 do not report the undefended AUROC, so it is unclear whether any method reduces the privacey leakage.
  2. For BotDrop's evaluation, some natural baselines are missing, e.g., leave-one-out on the surrogate, gradient saliency and attention rollout/flow.
Questions for authors
  1. What is the undefended AUROC for each attack and dataset for experiments of ATTENTIONSWAP?
  2. How does BotDrop compare with established token-importance methods including (i) leave-one-out deletion on the surrogate; (ii) gradient-based saliency; and (iii) attention rollout/flow?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.