Implementation:EvolvingLMMs Lab Lmms eval MMSearch Plus Decrypt Utils

Location: /tmp/kapso_repo_sslb_59s/lmms_eval/tasks/mmsearch_plus/decrypt_utils.py

Principle: Task_Utility_Functions

Purpose

Decryption utilities for MMSearch-Plus dataset that protects sensitive data using XOR cipher with password-derived keys.

Key Functions

derive_key

def derive_key(password: str, length: int) -> bytes

Derives encryption key from password:

• Uses SHA-256 hashing on password
• Repeats/truncates hash to match desired length
• Returns derived key bytes of specified length

Algorithm:

key = sha256(password)
result = key * (length // len(key)) + key[:length % len(key)]

decrypt_text

def decrypt_text(ciphertext_b64: str, password: str) -> str

Decrypts base64-encoded ciphertext:

• Returns original if input empty
• Decodes base64 to get encrypted bytes
• Derives key matching ciphertext length
• XOR decrypts: decrypted[i] = encrypted[i] ^ key[i]
• Decodes result as UTF-8
• Returns original on failure (logs warning)

decrypt_sample

def decrypt_sample(sample: Dict[str, Any], canary: str) -> Dict[str, Any]

Decrypts all encrypted fields in a dataset sample:

Encrypted Text Fields:

• question
• video_url
• arxiv_id

Encrypted List Field:

• answer (list of strings, each decrypted separately)

Unencrypted Fields:

• Images (NOT encrypted)
• category
• difficulty
• subtask

Returns decrypted copy of sample dictionary.

Implementation Details

• XOR cipher with SHA-256 derived keys
• Base64 encoding for encrypted data
• Canary string serves as password (e.g., "MMSearch-Plus")
• Graceful fallback: returns original text if decryption fails
• Selective field encryption for privacy protection
• List fields handled element-wise