Project description
AMD is building a hardware-assisted security platform that uses silicon-level Performance Monitoring Counters (PMCs) and on-chip machine learning to detect advanced endpoint threats (ransomware, fileless malware, cryptojacking) at the processor layer, below OS-based evasion. The platform collects CPU behavioral telemetry, classifies it via an ML inference engine, and exposes threat signals to security-software partners through a standardized API. The team covers the full stack: silicon telemetry, ML training/validation, real-time inference, lab qualification and CI/CD.
Responsibilities
- Design, train and evaluate ML classifiers (binary and multi-class) on CPU PMC telemetry datasets targeting ransomware, cryptomining, fileless malware and related threat categories.
- Perform feature engineering on raw hardware performance counter data (branch behavior, cache miss patterns, instruction mix ratios, execution port utilization) to extract discriminative threat signatures.
- Implement evaluation frameworks measuring detection rate, false-positive rate and inference latency on target GPU/NPU hardware.
- Expand and validate training datasets across malware variants; iteratively improve model coverage and accuracy.
- Optimize model architectures for inference on AMD integrated GPU and NPU hardware, balancing accuracy against strict CPU overhead targets. Export models to production-compatible inference formats and collaborate with real-time developers for pipeline integration.
- Document model architecture decisions, training-data provenance, evaluation metrics and known limitations.
- Maintain version control and reproducibility for all training pipelines and model artifacts.
Skills
Must have
- 3+ years of industry experience in applied ML or data science.
- Proficiency in Python; strong hands-on experience with PyTorch, TensorFlow or scikit-learn.
- Experience with binary or multi-class classification on tabular or time-series data. Solid understanding of model evaluation: cross-validation, precision/recall, F1, ROC-AUC.
- Understanding of model optimization for inference: quantization, pruning, ONNX export.
Nice to have
- Experience with anomaly detection or one-class classification methods. Background in cybersecurity, malware analysis or endpoint threat detection.
- Familiarity with hardware performance counters or systems-level telemetry as ML input features.
- Experience training models for deployment on GPU or NPU accelerators with constrained compute budgets.
Other
Languages: English: B2 Upper Intermediate
Seniority: Regular
