新型光學人工智慧能高精準度檢測深偽技術
New light-powered AI detects deepfakes with high accuracy
Updated at: October 2, 2026 at 04:45 AM
深偽技術變得愈發驚人地逼真,導致人們愈發難以辨別真相與虛構。目前,加州大學洛杉磯分校(UCLA)的研究人員取得了突破性進展:研發出一種光學神經處理器,利用光而非傳統運算來識別深偽內容。與傳統數位系統逐一處理影片幀的方式不同,該技術採用了一種混合架構。它將影片數據轉化為光模式,並使其穿過光學層。這使得處理器能夠在單次傳輸中同時分析超過15個影片串流。此方法不僅速度極快,而且能源效率極高。在測試中,該系統達到了97.79%的準確率,以99.86%的敏感度成功標記了幾乎所有偽造內容。此外,它還證明了對對抗性攻擊的韌性,意味著駭客很難欺騙該系統。隨著生成式人工智慧的不斷演進,這種以光為動力的技術提供了一種至關重要的高吞吐量防禦機制。透過將主要的檢測工作交由物理光學處理,它克服了傳統數位硬體面臨的能源與延遲問題,在保障我們的數位媒體環境方面邁出了重要一步。
Deepfakes are becoming alarmingly realistic, making it harder to distinguish between truth and fabrication. Now, researchers at UCLA have created a breakthrough: an optical-neural processor that uses light instead of traditional computing to catch deepfakes. Unlike digital systems that process video frames one by one, this technology uses a hybrid architecture. It converts video data into light patterns that travel through optical layers. This allows the processor to analyze over 15 video streams at the same time in a single pass. This method is not only incredibly fast but also highly energy-efficient. In testing, the system achieved a 97.79% accuracy rate, successfully flagging nearly all fake content with a sensitivity of 99.86%. Furthermore, it has proven resilient against adversarial attacks, meaning hackers struggle to trick the system. As generative AI continues to evolve, this light-powered technology offers a vital, high-throughput defense mechanism. By offloading the primary detection work to the physics of light, it overcomes the energy and latency issues faced by conventional digital hardware, marking a significant step forward in securing our digital media landscape.
