Workshops Archive

TMD - A Label-Free, Misalignment-Robust Distance for HPC Attack Detection


Workshop: 5th International Workshop on Cyber Security in High Performance Computing

Authors: Venkat Sai Suman Lamba Karanam (Bowling Green State University, Argonne National Laboratory); Zahmeeth Sakkaff (West Virginia University Institute of Technology, Argonne National Laboratory); and Shubham Sundriyal (Bowling Green State University)

Abstract: In this paper, we present the Temporal Mahalanobis Distance (TMD), a distortion-robust, label-free distance for attack detection in HPC networks. The telemetry from the distributed facility components in HPC ecosystem is temporally misaligned, since the components differ in sampling rate, in when an attack reaches them, and in sample ordering. Additionally, the operational monitoring is label-scarce. TMD makes the Mahalanobis distance separation-aware by scaling spatial deviations against the timestamp separation. TMD is a general distortion-robust distance, and we found it most valuable in the HPC regime, where the discriminative signal is the cross-node structure. TMD sustains the attack-normal separability under the offset and the ordering distortion, where Euclidean and DTW perform near chance. We lift TMD into a Sinkhorn distance, namely, TMSD, which helps under severe misalignment. As another contribution, we build a just-in-time streaming detector on a calibrated, semi-supervised TMD head that scores each window in sub-millisecond time.


Back to 5th International Workshop on Cyber Security in High Performance Computing Archive Listing

Back to Full Workshop Archive Listing