Privacy-Preserving News Recommendation over Homomorphic Encryption

Eishin Nakano, Takuya Suzuki, Yuki Yada, Hayato Yamana

Abstract: Online news platforms typically recommend news by evaluating the similarity between user preference embeddings and news content embeddings. However, such practices raise privacy concerns as platforms gain access to user preferences and recommendation results. Although homomorphic encryption (HE) offers a solution by enabling computation on encrypted data, it suffers from high latency and potential accuracy degradation due to approximation constraints. We propose a privacy-preserving and low-latency news recommendation method that ensures accuracy. The method operates in four phases: the platform server pre-calculates the news embeddings (phase 0); the user retrieves previously viewed news embeddings using HE (phase 1); the user computes the user embedding locally in plaintext (phase 2); and the platform server calculates encrypted click scores for candidate news using the encrypted user embedding (phase 3). This approach ensures user privacy by preventing the platform from accessing plaintext user embeddings and recommendation results. In our evaluation, the proposed method demonstrates practical latencies—312.8 ms for 1,000 stored news in phase 1, 2.5 ms in phase 2, and 433.8 ms for 50 candidate news in phase 3. Moreover, it maintains the same accuracy as when executing without encryption.