AI→Sec
Module 4: Side-Channel & ICS/IoT
Side-Channel, Hardware, and ICS/IoT Security
Learning from noisy physical data; securing OT & embedded systems
Outcome: Apply ML to power/EM/timing/RF streams and secure industrial/embedded environments.
Learning Objectives (3)
- Build CNN/Transformer-based pipelines for power/EM side-channel analysis
- Model multivariate ICS telemetry for anomaly/attack detection with explainability
- Evaluate RF/device fingerprinting and attestation under domain shift
Topic Map (5)
- CNN/Transformer SCA pipelines
- Masking/shielding/randomization
- ICS anomaly detection
- Device identity & attestation
- Wireless/RF fingerprinting
Topic Map — Deep Dives (5)
- CNN/Transformer SCA pipelines
- Acquisition: shunt vs. EM probe; sampling rate, trigger, desync sources.
- Preprocess: alignment (XCorr/DTW), denoise, normalization, POI selection.
- Models: 1D-CNN, residual stacks, Transformers; triplet/contrastive losses for templates.
- Augment: jitter, crop, mixup, noise; calibration for distribution shift.
- Metrics: success rate vs. #traces, guessing entropy, key rank.
- Masking/shielding/randomization
- Leakage models: HW/HD; higher-order attacks under masking.
- Random delay/instruction shuffling; clock jitter; hiding vs. masking trade-offs.
- Board/packaging effects; probe placement; regression to quantify SNR drops.
- ICS anomaly detection
- Signals: PLC tags (level/flow/pressure), setpoints, actuator states, alarms.
- Forecasting vs. reconstruction: ARIMA/VAR baselines → AEs/LSTM/Transformers.
- Context: maintenance windows, phase changes, and seasonality handling.
- Explainability: variable contributions; safe playbooks and operator-friendly alerts.
- Device identity & attestation
- Roots of trust and measured boot; attest evidence and freshness.
- PUFs: stability vs. reliability; helper data and ML modeling concerns.
- Operationalizing identity in fleets (rotation, revocation, supply-chain ties).
- Wireless/RF fingerprinting
- IQ features, CFO/phase transients, preamble shapes; packet-level vs. stream-level views.
- Deep sequence/CRF models; robustness to SNR/channel fading.
- Domain adaptation and calibration across environments (temp/aging).
Key Shifts Powered by AI (3)
- Profiling attacks at scale CNNs learn leakage features directly, reducing manual preprocessing and boosting success on real devices. [tches20_cnn] [ascad] Why it matters: Faster SCA prototyping and evaluation of countermeasures.
- From rules to learned ICS monitors Deep models capture multivariate process dynamics for anomaly/attack detection in ICS. [ndss24_ics] Why it matters: Earlier detection with fewer handcrafted thresholds.
- RF identities via deep sequence models Deep CRF/sequence models improve robustness of RF device fingerprinting. [tifs24_deepcrf] Why it matters: More reliable device ID without full crypto handshakes.
Still Hard (5)
- Generalization across devices, process phases, and environments.
- Leakage desynchronization and countermeasures reducing SNR.
- Data scarcity and labeling in ICS plants; safety constraints on data collection.
- RF domain shift (temperature/aging/channel) and spoofer adaptation.
- Privacy/regulatory limits for telemetry retention and cross-site model sharing.
References
- Zaid et al. “A Methodology for Efficient CNN Architectures in Profiling Attacks.” IACR TCHES 2020(1).
- Prouff et al. “Study of Deep Learning Techniques for Side-Channel Analysis and Introduction to ASCAD Database.” IACR ePrint 2018.
- Fung et al. “Attributions for ML-based ICS Anomaly Detection: From Theory to Practice.” NDSS 2024.
- DeepCRF: Deep Learning-Enhanced CSI-Based RF Fingerprinting for Channel-Resilient WiFi Device Identification