Secure Mobile AI for iOS and Android

• Supervised vs unsupervised learning: classification, regression, clustering
• Neural networks basics: architecture, forward & backward pass, activation, loss functions
• Generative AI: what it is, generative models vs discriminative; recent trends applicable to mobile (e.g. diffusion, lightweight LLMs)
• Performance metrics: accuracy, precision/recall, F1, latency, inference time, model size, energy cost
• Overfitting, underfitting, generalization; dataset quality issues (distribution shift, bias)

On-device, cloud and hybrid architectures

• Latency, bandwidth, offline behaviour, privacy, model size, operating cost and update pathways.
• LiteRT, the successor to TensorFlow Lite, and Core ML for model conversion and inference.
• ML Kit APIs by platform and feature; Android GenAI APIs use Gemini Nano through AICore on supported devices.
• Apple Foundation Models and other platform APIs: check system, model and device availability rather than assuming universal support.
• Cloud fallback or split processing where appropriate; evaluate the added transmission and availability requirements.

• Computer Vision
  • Image classification, object detection, segmentation
  • Use-cases: AR overlays, pose detection, OCR, face detection
  • SDKs / examples: Vision frameworks (iOS), ML Kit / TensorFlow Lite for Android, open models
• Natural Language Processing (NLP)
  • Tokenization, embeddings, simple transformer-based models
  • Tasks: text classification, summarization, translation, voice assistants, speech-to-text / text-to-speech
  • Consider on-device vs cloud approach depending on resource & privacy constraints
• Recommendation Engines & Personalization
  • Collaborative vs content-based vs hybrid recommendations
  • Collecting & processing user behaviour data in privacy preserving ways
  • Model freshness, feedback loops
• Generative Features (if time)
  • Lightweight generation: small LLMs, prompt-based (on-device or via API)
  • Use cases: autocomplete, suggestions, image style transfer

Build one selected feature from a prepared model or API; other use cases are demonstrations or design exercises.

• Set up suitable ML Kit, LiteRT or Core ML tooling for the chosen sample.
• Check supported model conversion paths from TensorFlow or PyTorch; unsupported operators may require adaptation.
• Profile latency, memory and energy on representative devices.
• Discuss quantisation, pruning and distillation as model-dependent techniques, not automatic quality-preserving switches.
• Evaluate acceleration, app-packaged versus downloaded models, versioning, integrity checks and recovery from failed updates.
• Data collection, feedback loops and model freshness; federated learning as a conceptual option.
• Model-IP and tampering risks: client-side protection reduces risk but cannot guarantee prevention of reverse engineering.

• Minimise collected data, use platform permissions and suitable consent, and protect local storage and network transmission.
• Assess demographic, device and edge-case performance; review distribution shift and bias mitigation.
• Provide meaningful user explanations and appropriate logs without exposing sensitive inputs.
• Evaluate adversarial inputs, model leakage, API misuse, tampering and update integrity.
• Identify applicable privacy and AI-law questions for qualified review; no universal compliance or certification claim.
• Review the sample integration and document security, resource and evaluation trade-offs.