Cepstral Analysis for Detection, Classification, Localization, and Mitigation of Cyber-Physical Attacks on AI Data Center Power

This record contains the complete code to reproduce every figure, table and number of the paper Cepstral Analysis for Detection, Classification, Localization, and Mitigation of Cyber-Physical Attacks on AI Data Center Power by Y. B. Tadese, M. Chamana, O. Adeyanju and S. Bayne (Texas Tech University), submitted to IEEE Transactions on Industrial Cyber-Physical Systems. The method turns two signals into power cepstra: AI data center power telemetry, and a physics-consistency residual against a trusted point-of-common-coupling meter. A random forest uses these together with gain-normalized spectral features to detect, classify and localize eight attacks: replay, time delay, square and sinusoidal malicious load oscillation, scaling, bias, ramp and masking. The estimated cepstral parameters then tune mitigation. This includes a period-selective synchronous-comb battery energy storage (BESS) compensator and a predictive form that compensates for converter delay and lag in realistic hardware. Contents: a dependency-free parquet reader; facility scenarios built from measured traces, with run-disjoint train/test/out-of-distribution splits; the eight attack models; cepstral, statistical, spectral (PSD) and autocorrelation (ACF) features; detection, classification, localization and zero-day experiments; delay and replay-loop estimation; telemetry restoration and BESS mitigation; robustness sweeps; stress tests and hardening: compromised meter, coordinated multi-feed and adaptive replay attacks, field telemetry impairments; a realistic BESS study; figure and table generation; the LaTeX sources of the paper and its supplementary material; a one-command reproduction script (reproduce_paper.py); a verifier (verify_results.py) that compares regenerated results with the reference values in results_reference/.