Course Objective & Scope
Training and SFT/LoRA fine-tuning pipelines fed by web scrapes, crowdsourced labels or user feedback loops allow attackers who control even a tiny fraction of samples to degrade model accuracy or implant a stealthy backdoor that activates only when a specific trigger appears. This course covers training-time data poisoning attacks and defensive engineering.
In hands-on labs participants execute BadNets, clean-label feature collision and LLM instruction-tuning poisoning attacks, then detect and remove poisoned samples and trojans using activation clustering, spectral signatures, influence functions and Neural Cleanse.
Who Should Attend?
- Data scientists and ML engineers responsible for model training and fine-tuning pipelines
- Data engineers managing data collection, annotation and data quality infrastructure
- AI security and model validation specialists
- Prerequisite: experience training or fine-tuning models in PyTorch and basic linear algebra
Key Takeaways
- Distinguish availability poisoning, targeted clean-label poisoning and backdoor/trojan attacks
- Implant a stealth trigger into a classifier or LoRA adapter using a small poisoning budget
- Filter poisoned training samples via latent-space activation clustering and spectral signature analysis
- Inspect and remediate trained models using Neural Cleanse trigger synthesis and fine-pruning
- Establish data provenance and quarantine gates for continual learning and feedback-driven pipelines
Curriculum
Taxonomy of Training-Time Poisoning Attacks
- Label flipping, bi-level optimisation and availability degradation attacks
- Clean-label feature collision and convex polytope targeted poisoning
Backdoor (Trojan) Attacks and Fine-Tuning Risks
- BadNets, imperceptible triggers and semantic backdoors
- Alignment degradation and sleeper-agent triggers in LLM SFT/LoRA and DPO/RLHF datasets
Dataset Sanitisation and Poison Detection
- Latent-space Spectral Signatures and Activation Clustering
- Identifying suspicious samples with k-NN label consistency and TracIn / influence functions
Model-Level Backdoor Inspection and Remediation
- Reverse-engineering triggers with Neural Cleanse and runtime detection with STRIP
- Erasing trojan pathways via fine-pruning, machine unlearning and distillation
Secure Data Ingestion and Feedback Loop Design
- Quarantine staging and golden-set regression gating before training on production user feedback
- Dataset versioning (DVC), cryptographic hashing and annotator reputation scoring
Other Courses in This Category
-
In-depth training on protecting vision, audio, biometric and multimodal deep neural networks against weight and architecture theft, feature reconstruction, sponge latency attacks and inference-layer compromises.
- Live Course
- 5 modules
-
Teaches how to secure the end-to-end MLOps lifecycle—across data pipelines, Jupyter workspaces, MLflow, Kubeflow, model registries, CI/CD and KServe/Kubernetes serving—using Sigstore/Cosign, ModelScan and Zero Trust controls.
- Live Course
- 5 modules
-
Covers the full OWASP Top 10 for LLM Applications—including direct and indirect prompt injection, improper output handling, unbounded consumption and sensitive data disclosure—from a secure software development (SSDLC) perspective.
- Live Course
- 5 modules
-
Hands-on training on conducting end-to-end penetration tests against classical ML, deep learning and LLM APIs using MITRE ATLAS, Counterfit, ART, garak and PyRIT, and producing actionable pentest reports.
- Live Course
- 5 modules
-
Hands-on training on applying machine learning and LLMs across cybersecurity—from phishing and malware detection to vulnerability prioritisation, cyber threat intelligence (CTI) analysis and SecOps assistants.
- Live Course
- 5 modules
-
Teaches how to detect malware, ransomware, C2 traffic and multi-stage attack chains across EDR, NDR and cloud telemetry using behavioural machine learning, sequence models and graph analytics.
- Live Course
- 5 modules
Get in touch
