RTAM-Fusion
A membership inference attack method targeting BPE tokenizers. It uses shadow tokenizers to calibrate the mass of rare tokens and fuses signals across rarity tiers.
Timeline (last 90 days)
The RTAM-Fusion method uses shadow tokenizers to calibrate the quality of rare tokens and fuses signals across rarity tiers.
On the tokenizer membership inference benchmark, RTAM-Fusion outperforms the Vocabulary Overlap baseline on both AUC and balanced accuracy while requiring fewer shadow tokenizers.
The Vocabulary Overlap baseline only counts the literal overlap between the candidate set and the vocabulary, ignoring frequency and rarity information, so RTAM-Fusion's improvement shows that paying attention to which rare tokens are retained is more discriminative than how many tokens overlap.
The code for RTAM-Fusion has been released in an anonymous repository for reproduction and further verification.
All Facts (4)
The RTAM-Fusion method uses shadow tokenizers to calibrate the quality of rare tokens and fuses signals across rarity tiers.
50%UnverifiedOn the tokenizer membership inference benchmark, RTAM-Fusion outperforms the Vocabulary Overlap baseline on both AUC and balanced accuracy while requiring fewer shadow tokenizers.
50%UnverifiedThe Vocabulary Overlap baseline only counts the literal overlap between the candidate set and the vocabulary, ignoring frequency and rarity information, so RTAM-Fusion's improvement shows that paying attention to which rare tokens are retained is more discriminative than how many tokens overlap.
50%UnverifiedThe code for RTAM-Fusion has been released in an anonymous repository for reproduction and further verification.
50%