Model Online / On-Site
Duration
Modules 5 modules

Course Objective & Scope

Deep neural networks powering computer vision, voice/biometric verification and multimodal (vision-language) workloads face critical threats when deployed to edge devices, on-prem GPU nodes or Triton/ONNX Runtime clusters—including weight theft, weight tampering (fault/bit-flip attacks) and reconstructing raw images or audio from intermediate embeddings.

This course covers architecture-specific vulnerabilities in CNNs and Transformers, latency-amplifying sponge examples, deep model watermarking, encrypted ONNX/TensorRT deployment and protecting model weights in GPU memory with NVIDIA Confidential Computing.

Who Should Attend?

  • Engineers developing computer vision, speech, biometric and multimodal deep learning models
  • Edge AI, embedded inference and high-performance GPU serving (Triton/TensorRT) teams
  • Security specialists responsible for protecting deep learning IP and model weights
  • Prerequisite: experience with deep neural architectures (CNNs/Transformers) in PyTorch and ONNX export

Key Takeaways

  • Analyse layer-by-layer confidentiality and integrity risks in CNNs, Vision Transformers and multimodal networks
  • Measure and mitigate raw input reconstruction from embeddings and split-inference activations
  • Protect inference servers against energy- and latency-amplifying sponge examples
  • Embed and verify weight-space and trigger-set watermarks to prove model ownership
  • Secure model weights at rest and in use with ONNX/TensorRT encryption, Triton hardening and Confidential GPUs

Curriculum

Attack Surfaces Across Deep Neural Architectures
  • Comparing vulnerability profiles of CNNs, ResNets, Vision Transformers (ViTs) and multimodal VLMs
  • Presentation/deepfake spoofing vs adversarial perturbations in biometric (face/voice) verification
Intermediate Layer and Embedding Reconstruction
  • Reconstructing raw images or text from feature embeddings and split-inference activations
  • Defences using Information Bottleneck regularisation, activation noise and dimensionality projection
Weight Integrity, Bit-Flip and Sponge Attacks
  • Targeted weight tampering and bit-flip attacks on quantized neural networks
  • Mitigating sponge examples that maximise energy consumption and inference latency
Deep Neural Network IP Protection and Watermarking
  • White-box (weight distribution / passport layers) and black-box (trigger-set) watermarking schemes
  • Testing watermark survival under fine-tuning, weight pruning and quantization
Secure Model Deployment: ONNX, Triton and Confidential GPUs
  • Encrypting and signing ONNX/TensorRT artefacts and hardening NVIDIA Triton Inference Server
  • Protecting weights in memory across edge and cloud using Confidential Computing (GPU TEEs)

Other Courses in This Category

  • MLOps Security and DevSecOps for Machine Learning Pipelines

    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.

    Model Security & AI in Cybersecurity Corporate
    Live Course
    5 modules
    View
  • LLM Application Security: OWASP Top 10 and Prompt Injection Defense

    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.

    Model Security & AI in Cybersecurity Corporate
    Live Course
    5 modules
    View
  • AI Red Teaming and Machine Learning Model Penetration Testing

    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.

    Model Security & AI in Cybersecurity Corporate
    Live Course
    5 modules
    View
  • AI Applications in Cybersecurity Operations and Defense

    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.

    Model Security & AI in Cybersecurity Corporate
    Live Course
    5 modules
    View
  • AI-Driven Cyber Threat Detection for Endpoint, Network and Cloud

    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.

    Model Security & AI in Cybersecurity Corporate
    Live Course
    5 modules
    View
  • Anomaly Detection with Machine Learning for Security and Operations

    Teaches how to build unsupervised and semi-supervised anomaly detection pipelines on logs, network traffic and user behaviour (UEBA) using Isolation Forests, Autoencoders, time-series and graph methods while minimising false positives.

    Model Security & AI in Cybersecurity Corporate
    Live Course
    5 modules
    View

Get in touch