Selective ROI Image Encryption Pipeline

An end-to-end Python project that detects regions of interest (ROI) in images, encrypts only those regions using modern cryptography, and reconstructs the image after decryption for correctness verification.

Built for the Cryptography course.

---

Overview

Instead of encrypting an entire image, this pipeline:

1. Detects ROI areas (contour-based or face-based).
2. Extracts ROI pixels only.
3. Generates or derives an AES-256 key:
   • RSA mode (default): generates a random AES key and wraps it with the recipient's RSA public key.
   • Password mode: derives a key from a password using PBKDF2-HMAC-SHA256.
4. Encrypts ROI bytes using AES-256-GCM (default) or AES-256-CBC + HMAC-SHA256.
5. Signs metadata with the encryptor's RSA private key (RSA mode) for non-repudiation.
6. Stores encrypted payload + metadata (mask, shape, encrypted AES key or salt, etc.).
7. Verifies digital signature, decrypts, and reconstructs the original image.
8. Reports performance and security metrics (NPCR/UACI, timing, ROI ratio, etc.).

---

Key Features

• Selective encryption: encrypts only ROI pixels, reducing encryption workload.
• Hybrid RSA + AES encryption: RSA-OAEP wraps the AES session key (no shared password needed).
• Digital signatures: RSA-PKCS1v15 signs metadata for non-repudiation.
• Password mode: PBKDF2-HMAC-SHA256 key derivation available as fallback.
• Multiple ROI methods: global, adaptive, otsu, face.
• Face detection fallback: when no faces are detected, the entire image is encrypted as a safety net.
• Authenticated encryption: AES-GCM with optional AAD binding, or AES-CBC with Encrypt-then-MAC.
• Correct NPCR/UACI metrics: measured against actual AES ciphertext, not visual placeholders.
• Built-in self-tests in key_manager.py and encryption.py.

---

Project Structure

File	Purpose
main.py	Full ROI encryption pipeline CLI and orchestration
roi_detector.py	ROI detection, thresholding, contouring, face detection
key_manager.py	Salt generation, PBKDF2 key derivation, KDF parameter handling
encryption.py	AES-GCM / AES-CBC+HMAC encryption/decryption + packet helpers
rsa_manager.py	RSA key pair generation, OAEP key wrapping, digital signatures
keys/	Generated RSA key pair (auto-created on first run)
outputs/	Generated outputs (masks, encrypted image, metadata, etc.)

---

Requirements

• Python 3.9+
• Packages: numpy, opencv-python, matplotlib, pycryptodome

pip install numpy opencv-python matplotlib pycryptodome

---

Quick Start

RSA mode (default)

# First run auto-generates RSA keys in keys/
python main.py --image photo.jpg

Generate keys explicitly

python main.py --generate-keys

Password mode (legacy)

python main.py --image photo.jpg --key-mode password --password "StrongPass@123"

Other options

# Use face detection with CBC mode
python main.py --image photo.jpg --method face --mode CBC

# Change visual fill for encrypted ROI
python main.py --image photo.jpg --fill blur

# Run wrong-password rejection demo (password mode)
python main.py --image photo.jpg --key-mode password --wrong-password-test

# Speed vs security tradeoff table
python main.py --image photo.jpg --key-mode password --iteration-sweep

---

CLI Arguments

Argument	Default	Description
--image / -i	demo images	Input image path(s)
--key-mode	rsa	rsa (hybrid encryption) or password (PBKDF2)
--public-key	keys/public.pem	RSA public key PEM path
--private-key	keys/private.pem	RSA private key PEM path
--generate-keys	-	Generate RSA key pair and exit
--password / -p	-	Encryption password (password mode only)
--method / -m	otsu	ROI method: global, adaptive, otsu, face
--mode	GCM	AES mode: GCM or CBC
--iterations / -n	100000	PBKDF2 iterations (password mode only)
--fill	black	Visual fill: black, blur, noise
--min-area	500	Minimum contour area for ROI filtering
--output / -o	outputs	Output directory

---

How the Two Key Modes Work

RSA Mode (default)

Encryption:
  1. Generate random 32-byte AES key
  2. Encrypt AES key with recipient's RSA public key (RSA-OAEP)
  3. Encrypt ROI with AES key
  4. Sign metadata with sender's RSA private key
  5. Store: encrypted AES key + ciphertext + signature

Decryption:
  1. Verify metadata signature with sender's public key
  2. Decrypt AES key with recipient's RSA private key
  3. Decrypt ROI with AES key

This is the same approach used by TLS, PGP, and S/MIME.

Password Mode

Encryption:
  1. Generate random salt
  2. Derive AES key from password via PBKDF2-HMAC-SHA256
  3. Encrypt ROI with AES key
  4. Store: salt + iterations + ciphertext

Decryption:
  1. Re-derive AES key from same password + stored salt
  2. Decrypt ROI

---

Metrics

The pipeline reports both crypto and visual NPCR/UACI:

• NPCR/UACI (crypto): measured against actual AES ciphertext bytes interpreted as pixel values. This is the correct measurement for evaluating cipher quality.
• NPCR/UACI (visual): measured against the display fill (black/blur/noise). This is what you see on screen but does not reflect cipher quality.

---

Security Notes

• RSA mode avoids the shared-password problem entirely.
• Digital signatures provide non-repudiation (proof of who encrypted).
• Face detection falls back to full-image encryption when no faces are found, preventing accidental data exposure.
• Prefer AES-GCM whenever possible.
• Keep private keys secure. The keys/ directory is for learning only.

Warning: The default keys/ directory stores private keys in plaintext on disk. In a production system, use a hardware security module (HSM) or a key management service. This project is for educational use.

---

Production Considerations

This project is a working academic demonstration. For production deployment, the following integrations would be appropriate:

X.509 Certificates

The RSA public keys could be wrapped in X.509 certificates, enabling identity verification through a Certificate Authority (CA) trust chain. This is how HTTPS establishes identity - the browser trusts a CA, the CA vouches for the server's public key. For this project, each user's public key would be bound to their identity via a signed certificate.

JWT for Metadata

The metadata JSON could be packaged as a signed JWT (JSON Web Token) with standard claims (iss, iat, exp, sub). This would make the metadata interoperable with any JWT library in any language, and the exp claim could enforce time-limited access to encrypted images.

Cloud KMS Integration

The generate_key_pair() and key storage functions could be replaced with calls to AWS KMS, Azure Key Vault, or GCP Cloud KMS. These services provide hardware-backed key storage, automatic key rotation, fine-grained access control, and audit logs - none of which a local keys/ directory can offer.

Modern Face Detection

The Haar cascade face detector (Viola-Jones, 2001) could be replaced with MTCNN, RetinaFace, or MediaPipe for higher accuracy and per-detection confidence scores. The fallback-to-full-image mechanism would then trigger based on confidence thresholds rather than binary presence/absence.

---

Limitations

• ROI quality depends on image characteristics and chosen threshold method.
• Haar cascade face detection has limited accuracy on non-frontal faces.
• UACI values may deviate from ideal 33.5% depending on image content and ROI size, since AES ciphertext statistics depend on plaintext distribution.
• This is an educational pipeline, not a hardened production system.

---