Assignment Instructions Summary (Track B: Membership Inference)

This assignment is based on the official assignment brief document and the materials covered in Weeks 14 lectures and tutorials. The main objective is to perform a security and privacy review of an AI service, focusing on identifying vulnerabilities, testing them practicly, and proposing suitable defences.

The analysis must follow the structure provided in the assignment file, including:

• System understanding (architecture and trust boundaries)
• Identification of assets, attackers, and risks
• Designing an attack workflow
• Proposing defence strategies with justification
• Supporting analysis with figures, tables, and experimental evidence

  I will choos Track B: Membership Inference.This track focuses on analysing whether an attacker can determine if a specific data record was used in training the model, based on the models outputs.The implementation will be based on the provided file:

  • membership_inference_track.py
    

    Report Requirements

    The final submission will be:

    • One PDF report only
    • No code files will be submitted

    The report must:

    • Be scenario-specific (AI hiring API)
    • Include:
      • Figures (system architecture, attack workflow, defence workflow)
      • Tables (assetattackerharm)
      • Screenshots from code execution as evidence
    • Demonstrate:
      • Understanding of the system
      • Identification of vulnerabilities
      • Clear attack methodology
      • Justified defence design with trade-offs
        

        Coding Task (Track B)

        The coding task is minimal and only requires completing two missing lines in the provided script, After completing these lines, the script must be executed to generate results for analysis

        Required Screenshots

        Only screenshots are required (NOT full code submission). The report must include:

        1. Code Screenshot

        • Show ONLY the modified lines (the two completed lines in run_mia())

        2. Output Screenshot

        • Show the program output after running the code, including:
          • Baseline case results
          • Defended case results
        • Important values to capture:
          • member_mean_signal
          • nonmember_mean_signal
          • mia_attack_accuracy
          • threshold

        These outputs will be used as evidence in your discussion.

        Expected Analysis (Track B)

        In the report, you will:

        • Compare baseline vs defended model
        • Explain how defence techniques (e.g. regularization, label smoothing, dropout) reduce membership inference risk
        • Use your code results as supporting evidence

          please reffer to the attached files and read the assignment instructions careflly

    • • After you finished, provied a (Turnitin + AI) reports. Al