In brief

  • Novel approaches to differential privacy in offline reinforcement learning and code generation.
  • LLM agents continue to be targeted via prompt injection on their tool selection mechanisms.
  • LLMs are also increasingly leveraged for formal verification of smart contracts and protocol compliance.

PrivCode: When Code Generation Meets Differential Privacy

  • Fine-tuning Large Language Models (LLMs) on private code repositories risks leaking sensitive or proprietary data.
  • The authors propose PrivCode, a differentially private synthesizer for code datasets with strict syntactic dependencies.
  • It employs a two-stage approach: “privacy-sanitizing” with DP-SGD that includes syntactic information, followed by “utility-boosting” where a pre-trained LLM is fine-tuned on the synthetic code to recover utility.
  • Tested across four LLMs, the framework reportedly generates higher-utility code for various tasks under four benchmarks while maintaining privacy bounds.

Zheng Liu, Chen Gong, Terry Yue Zhuo, Kecen Li, Weichen Yu, Matt Fredrikson, Tianhao Wang. “PrivCode: When Code Generation Meets Differential Privacy.” NDSS — https://www.ndss-symposium.org/ndss-paper/privcode-when-code-generation-meets-differential-privacy/

Prompt Injection Attack to Tool Selection in LLM Agents

  • LLM agents frequently use a retrieval-and-selection pipeline to choose appropriate tools for a given task.
  • ToolHijacker is a prompt injection attack targeting this selection step in no-box scenarios.
  • The attack injects a malicious document into the tool library designed to consistently compel the LLM agent to select the attacker’s tool for a target task.
  • The authors formulated the malicious document creation as an optimization problem, bypassing both prevention-based (StruQ, SecAlign) and detection-based defenses.

Jiawen Shi, Zenghui Yuan, Guiyao Tie, Pan Zhou, Neil Zhenqiang Gong, Lichao Sun. “Prompt Injection Attack to Tool Selection in LLM Agents.” NDSS — https://www.ndss-symposium.org/ndss-paper/prompt-injection-attack-to-tool-selection-in-llm-agents/

ProtocolGuard: Detecting Protocol Non-compliance Bugs via LLM-guided Static Analysis and Dynamic Verification

  • Natural-language protocol specifications are ambiguous, often leading to non-compliance bugs in network implementations.
  • ProtocolGuard combines LLM-guided static analysis with dynamic verification to identify these bugs.
  • It extracts normative rules, performs program slicing, and detects semantic inconsistencies. It then automatically generates assertions and test cases for dynamic verification.
  • Evaluated on 11 protocol implementations, the framework discovered 158 non-compliance bugs, 70 of which were confirmed.

Xiangpu Song, Longjia Pei, Jianliang Wu, Yingpei Zeng, Gaoshuo He, Chaoshun Zuo, Xiaofeng Liu, Qingchuan Zhao, Shanqing Guo. “ProtocolGuard: Detecting Protocol Non-compliance Bugs via LLM-guided Static Analysis and Dynamic Verification.” NDSS — https://www.ndss-symposium.org/ndss-paper/protocolguard-detecting-protocol-non-compliance-bugs-via-llm-guided-static-analysis-and-dynamic-verification/

Also published