Out-of-distribution Detection in Dependent Data for Cyber-physical Systems with Conformal Guarantees

• DOI: 10.1145/3648005
• Publication Date: 2024-02-13T13:51:36Z
• Authors: Ramneet Kaur; Yahan Yang; Oleg Sokolsky; Insup Lee
• Keywords: Sliding window protocol; TRACE (psycholinguistics); False alarm

Uncertainty in the predictions of learning-enabled components hinders their deployment safety-critical cyber-physical systems (CPS). A shift from training distribution a component (LEC) is one source uncertainty LEC’s predictions. Detection this or out-of-distribution (OOD) detection on individual datapoints has therefore gained attention recently. But many applications, inputs to CPS form temporal sequence. Existing techniques for OOD time-series data either do not exploit relationships sequence provide any guarantees detection. We propose using deviation in-distribution equivariance as non-conformity measure conformal anomaly framework CPS. Computing independent multiple detectors based proposed and combining these by Fisher’s method leads detector CODiT with bounded false alarms. performs fixed-length windows consecutive Fisher value input window. further performing real-time traces variable lengths This can be done compute values sliding trace merging function. Merging functions such Harmonic Mean, Arithmetic Geometric Bonferroni Method, so on, used combine trace, combined alarm rate guarantees. illustrate efficacy achieving state-of-the-art results two case studies windows. The first an autonomous driving system perception (or vision) LEC. second study medical walking pattern GAIT analysis where physiological (non-vision) collected force-sensitive resistors attached subject’s body. For length traces, we consider same analysis. report our four computed trace. also compare study. Code, data, trained models are available at https://github.com/kaustubhsridhar/time-series-OOD .