使用混合位元長度乘法的安全訓練

Abstract

機器學習在應用上的快速發展引發了人們對資料隱私的擔憂，促使人們對於隱私維護機器學習的研究。為了確保機器學習訓練資料的隱私，一個有潛力的解決方案是使用涉及高成本密碼技術的雙方安全計算。 然而，這導致這種訓練方式的效率比明文的方式低幾個數量級。 矩陣乘法在基於雙方安全方計算的隱私保護機器學習中通常會消耗大量時間，我們利用混合位元長度乘法來取代固定位元長度的通用乘法，以提高矩陣乘法的效率。 另外，我們也提供調整位元長度以防止溢出的機制。 我們透過利用預先存在的資訊和具有較小誤差的高效率截斷來提高針對訓練模型的對數效率。我們的實驗表明，與使用相同的函式庫的固定位元長度訓練相比，我們的混合位元長度訓練減少了至少 22% 的時間延遲，同時減少了至少 34% 的通訊開銷。

The rapid development of machine learning applications has raised concerns about data privacy, driving the need for privacy-preserving machine learning research. To ensure the privacy of machine learning training data, a potential solution is to use secure two-party computation involving high-cost cryptography techniques. However, this results in training efficiency that is several orders of magnitude lower than plaintext machine learning. Our work utilizes multiplication with mixed bit-length to replace generic multiplication with fixed bit-length to enhance the efficiency of matrix multiplication, which typically consumes a significant amount of time in privacy-preserving machine learning based on secure two-party computation. We also provide mechanisms that adjust the bit-length to prevent overflow. In addition, we improve the efficiency of the logarithm for the training model by leveraging pre-existing information and efficient truncation with a minor error. Our experiment demonstrates that our mixed bit-length training has at least 22% latency reduction along with at least 34% communication overhead reduction compared to fixed bit-length training using the same two-party computation library.