學習針對即將發佈的影像資料之隱私保護嵌入

Abstract

在現今的研究中，深度學習已成為一種強而有力的技術，並且在許多問題上取得重大的進展。卷積神經網路與大量的影像資料使得影像處理領域的研究快速而蓬勃地發展。然而，當人們使用深度學習的技術來解決問題時，不可避免得會遇到隱私洩露的問題。而隱私全為基本人權之一，故隱私洩漏的問題也成為我們需要攻克的議題。由於使用大規模的影像資料，使得對於原始資料與隱藏於影像中的敏感訊息的隱私洩漏問題成為必須關注的議題。因此，特別是對肉眼不好辨識的影像，例如：隱藏有性別資訊的Ｘ光圖片，我們提倡釋放出隱私保護的嵌入以取代釋出原始影像。除了避免使用者直接接觸原始影像，使用嵌入還可以用以避免用原始影像導致的特定敏感資訊的隱私洩漏風險。 為了達成這樣的目的，在採用不同方法進行實驗後，「 混合」模型最終被我們採用。「 混合」是多目標學習模型，它採用分解特徵的概念作為核心技術，加上以特定方法先訓練出初始權重，並且以對抗示例圖作為訓練輸入。多目標網路以底層的分享層作為特徵擷取器和兩個分別解決主任務與輔助任務的辨別器組合而成。特徵擷取器和輔助任務辨別器進行對抗過程來優化輔助任務。我們在MAFL人臉資料集和NIH提供的胸腔Ｘ光圖資料集進行實驗。 其結果展露出由混合模型擷取器生成的嵌入可以成功地預測主要任務且可以將指定的敏感資訊去除。更甚之，我們發現將混合模型得來的嵌入加上叉分隱私技術可以得到更好的表現。

Deep learning is a powerful technique which make a great process in solving different problems. The usage of convolutional neural network and massive image data let researches about image processing develop rapidly. However, when deep learning is utilized, the problems about privacy leakage need to be concerned simultaneously. Due to large-scale image data, privacy preserving for original data and sensitive information hidden on it is essential. Therefore, we purpose releasing privacy-preserving embeddings, especially for images which sensitive information is invisible to the naked eye, e.g., X-ray images with gender information hidden behind, replacing to original image data. The embeddings are able to avoid privacy leakage of original image data and specific sensitive information. To reach our goal, after conducting several methods, hybrid model is purposed finally. Hybrid is a multitask-learning model for disentangling features with good initial weights by iterative training and adversarial examples as inputs. The multitask network composes of some shared layers on the bottom as a feature extractor and two discriminators respectively for a main task and a sensitive task. The feature extractor and the sensitive discriminator conduct an adversarial process to optimize sensitive loss. The experiments on fatial database MAFL and medical image database NIH Chest X-ray demonstrate that embeddings generated by the hybrid extractor can predict the main task with designated sensitive information being wiped out. Moreover, it is discovered that hybrid model with differential privacy leads to a better performance.