IF-CPS: Influence Functions for Cyber-Physical Systems -- A Unified Framework for Diagnosis, Curation, and Safety Attribution
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866912978386812928 |
|---|---|
| author | Li, Jiachen Li, Shihao Bakshi, Soovadeep Xu, Jiamin Chen, Dongmei |
| author_facet | Li, Jiachen Li, Shihao Bakshi, Soovadeep Xu, Jiamin Chen, Dongmei |
| contents | Neural network controllers trained via behavior cloning are increasingly deployed in cyber-physical systems (CPS), yet practitioners lack tools to trace controller failures back to training data. Existing data attribution methods assume i.i.d.\ data and standard loss targets, ignoring CPS-specific properties: closed-loop dynamics, safety constraints, and temporal trajectory structure. We propose IF-CPS, a modular influence function framework with three CPS-adapted variants: safety influence (attributing constraint violations), trajectory influence (temporal discounting over trajectories), and propagated influence (tracing effects through plant dynamics). We evaluate IF-CPS on six benchmarks across diagnosis, curation, and safety attribution tasks. IF-CPS improves over standard influence functions in the majority of settings, achieving AUROC $1.00$ in Pendulum (5-10\% poisoning), $0.92$ vs.\ $0.50$ in HVAC (10\%), and the strongest constraint-boundary correlation (Spearman $ρ= 0.55$ in Pendulum). |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_21543 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | IF-CPS: Influence Functions for Cyber-Physical Systems -- A Unified Framework for Diagnosis, Curation, and Safety Attribution Li, Jiachen Li, Shihao Bakshi, Soovadeep Xu, Jiamin Chen, Dongmei Systems and Control Neural network controllers trained via behavior cloning are increasingly deployed in cyber-physical systems (CPS), yet practitioners lack tools to trace controller failures back to training data. Existing data attribution methods assume i.i.d.\ data and standard loss targets, ignoring CPS-specific properties: closed-loop dynamics, safety constraints, and temporal trajectory structure. We propose IF-CPS, a modular influence function framework with three CPS-adapted variants: safety influence (attributing constraint violations), trajectory influence (temporal discounting over trajectories), and propagated influence (tracing effects through plant dynamics). We evaluate IF-CPS on six benchmarks across diagnosis, curation, and safety attribution tasks. IF-CPS improves over standard influence functions in the majority of settings, achieving AUROC $1.00$ in Pendulum (5-10\% poisoning), $0.92$ vs.\ $0.50$ in HVAC (10\%), and the strongest constraint-boundary correlation (Spearman $ρ= 0.55$ in Pendulum). |
| title | IF-CPS: Influence Functions for Cyber-Physical Systems -- A Unified Framework for Diagnosis, Curation, and Safety Attribution |
| topic | Systems and Control |
| url | https://arxiv.org/abs/2603.21543 |