报告详情
SentinelDB-X Causal Temporal Runtime Verification for Cumulative Semantic Attacks in LLM Database Agents
编号：128
访问权限：仅限参会人
更新：2026-11-10 16:51:17 浏览：15次
Online
报告开始：2026年11月20日 15:15（Asia/Kolkata）
报告时间：15min
所在会场：[S7] Networking, Security and Blockchain Systems [S7-1] Networking, Security and Blockchain Systems
暂无文件
摘要
LLM-driven database agents are able to build individually admissible queries, which breach disclosure, purpose, provenance, or delegated authority in a session. We introduce SentinelDB-X, a runtime verifier that switches the target of enforcement from a single query to a causal temporal execution trace. The verifier keeps a typed Causal Temporal Exposure Ledger, generates intent to executable policy state, preserves authority monotonicity across retrieved data and tools, and applies bounded counterfactual replay before high impact actions. Neural SQL risk and open world evidence support the decision but not for execution. In all source held out, unseen family, generated attack, tool mediated, and cumulative session regimes, SentinelDB-X achieves an overall F1 score of 92.6 percent, compared to 67.8 percent for the highest baseline F1 score at the query level. The cumulative F1 is reduced to 73.8 percent without temporal memory. The full system attains 0.92 Top 3 causal event recall, 0.87 attribution nDCG, and no hard policy violations on the encoded policy test set, while the median latency is 10.6ms on the local path and 18.4ms on the counterfactual path.
关键词
LLM database agents,,Cumulative Semantic Attacks,Runtime Verification,Data Provenance,Counterfactual Replay,Neuro-Symbolic Security
报告人
FIRAS TARIK JASIM
Asst. Lect shivaji university稿件作者
全部评论
重要日期
-
会议日期
11月20日
2026
至11月22日
2026
-
10月31日 2026
初稿截稿日期
-
11月07日 2026
提前注册日期
-
11月22日 2026
注册截止日期
主办单位
United Societies of Science
承办单位
Pondicherry University
协办单位
IEEE Section
IEEE Madras Section
IEEE Madras Section
联系方式
- ICon INDIA 2026
- ic******@usssociety.org
发表评论