docs/threat_model.md

SecureLens — Threat Model Documentation

System Overview

SecureLens is a privacy-preserving medical image diagnostic system. A client uploads a chest X-ray, features are encrypted using CKKS Fully Homomorphic Encryption, and the cloud server runs inference on ciphertexts without accessing plaintext data.

---

Assets Being Protected

Asset	Sensitivity	Protection Method
Raw X-ray pixels	Critical	Never transmitted
512-dim feature vector	High	CKKS encrypted
Diagnostic result	Medium	Decrypted client-side
CKKS secret key	Critical	Client-side only
Model weights	Medium	Server-side plaintext

---

Threat Actors

Actor	Capability	Goal
Malicious cloud provider	Full server access	Read patient data
Network attacker (MITM)	Intercept traffic	Steal X-ray images
Server-side attacker	Read server memory	Access patient records
Insider threat	Physical server access	View diagnostic data

---

Attack Vectors and Mitigations

1. Server Data Breach

• Attack: Attacker gains full access to cloud server
• Without FHE: All patient X-rays exposed
• With FHE: Only CKKS ciphertexts stored — mathematically useless without secret key
• Status: MITIGATED

2. Network Interception (MITM)

• Attack: Attacker intercepts HTTP traffic
• Without FHE: Raw feature vectors readable
• With FHE: 326KB ciphertext — indistinguishable from random noise
• Status: MITIGATED

3. Image Tampering Attack

• Attack: Attacker modifies image during transmission
• Without FHE: Corrupted image reaches server → wrong diagnosis
• With FHE: Image encrypted before transmission → tampering affects only ciphertext noise
• Status: MITIGATED (demonstrated in attack demo)

4. Inference Attack on Model Output

• Attack: Attacker learns patient features from repeated queries
• Mitigation: CKKS noise + rate limiting on API
• Status: PARTIALLY MITIGATED

5. Timing Side-Channel

• Attack: Measure inference time to infer input characteristics
• Mitigation: Consistent O(n) linear inference time
• Status: LOW RISK

---

What FHE Does NOT Protect Against

1. Compromised client device (secret key stolen)
2. Malicious model weights (server-side poisoning)
3. Volume analysis (attacker sees number of queries)
4. Social engineering attacks

---

CKKS Security Parameters

Parameter	Value	Security Guarantee
Scheme	CKKS	Approximate HE for real numbers
Poly modulus degree	8192	128-bit security level
Coefficient modulus	[60,40,40,60]	3 multiplication levels
Global scale	2^40	~10 decimal digits precision
Security level	128-bit	Equivalent to AES-128
Decryption error	~7×10⁻⁸	Negligible for classification

---

Compliance

• HIPAA: Compliant (no PHI stored on server)
• India DPDP Act 2023: Compliant (sensitive health data encrypted)
• GDPR Article 25: Compliant (privacy by design)