Red-Team an Image-Classification Pipeline for a Banking KYC Workflow
The situation
The bank (around 9,000 staff, around 5 million retail customers in Ireland and the UK) processes about 12,000 KYC checks per day; a public-grade adversarial failure would trigger a regulatory disclosure to the Central Bank.
Your task
Quantify the production KYC image classifier's robustness to standard adversarial attacks and recommend three remediations.
The brief
You start with
You receive the production image classifier as a black-box API plus a labeled validation set of 5,000 ID images.
What you do
- Run untargeted FGSM and PGD attacks (L_inf budget 4/255 and 8/255) and a transfer attack from a surrogate model.
- Quantify robust accuracy and identify the top 3 attack patterns the model is most vulnerable to.
You deliver
The attack notebooks, a robust-accuracy table, a failure-mode gallery, and a 4-page risk memo with three concrete remediation recommendations.
What you will practice
- Implement and apply standard adversarial attacks to a real classifier
- Quantify robust accuracy at standard threat-model budgets
- Identify failure patterns and translate them into remediation plans
- Communicate adversarial-robustness findings to a non-research risk committee
A suggested way in
- Use torchattacks or foolbox for reference attack implementations
- Always report robust accuracy at the standard 4/255 and 8/255 L_inf budgets
- Transfer attacks from a surrogate often hurt more than direct PGD — include them
Skills you will practice
Adversarial attacks, Robust evaluation, Red teaming, Threat modeling, Pytorch, Risk assessment
Where it leads
AI Safety Researcher, ML Researcher, Applied AI Scientist
How it is assessed
- Attack methodology30%
Attacks implemented to standard specs at the right budgets.
- Robustness measurement25%
Robust accuracy reported with CIs and per-attack breakdown.
- Failure-mode insight20%
Gallery surfaces patterns, not just samples.
- Risk-memo quality25%
Risk committee can act on the three remediations.
Resources
- PyTorchTool
Underlying framework for attacks against the classifier.
- torchattacksTool
Reference Python library for adversarial attacks.
- PythonTool
Glue code for the attack harness.
- MatplotlibTool
Failure-mode visualizations.
- JupyterTool
Iterate on attack hyperparameters.
- NotionTool
Compose the 4-page risk memo.
Finish it. Earn this.
A verified certificate with your name. Anyone can check it.
Time to practice.
Solve this challengeMore like this.
- Applied AI ScientistFine-Tune a Vision-Language Model for Image Captioning
- Research ScientistEmbodied Visual Reasoning for a Warehouse Pick Assistant
- Product ManagerOptimizing Onboarding for a Fintech App
- Market Research AnalystBusiness Cycle Positioning for a Renewable Energy Firm
-
Fine-Tune a Vision-Language Model for Image Captioning Applied AI Scientist Take BLIP-2 or LLaVA-1.6 as the base. Fine-tune (LoRA is fine) on a 4,000-image accessibility-curated dataset where each image has a useful caption wr vision language modelslora fine tuningpytorch Open challenge → -
Embodied Visual Reasoning for a Warehouse Pick Assistant Research Scientist Use an embodied simulator (Habitat 3.0 or Isaac Sim — pick one and justify) to render 300 cluttered-bin scenarios with a target item label. For each s embodied visionvision language modelsvisual reasoning Open challenge → -
Optimizing Onboarding for a Fintech App Product Manager Your task is to analyze the current onboarding flow, identify drop-off points, and propose a redesigned onboarding experience. You will use user behav user researchux designdata analysis Open challenge → -
Business Cycle Positioning for a Renewable Energy Firm Market Research Analyst Your task is to analyze current business cycle indicators for Italy (GDP growth, unemployment, consumer confidence, interest rates) and the specific i business cycle analysisleading indicatorsdemand forecasting Open challenge →