- 期刊
- OpenAccess
基於AI語音特徵分析的智慧型手機深偽詐騙偵測研究
Research on AI-based Audio Feature Analysis for Deepfake Fraud Detection in Mobile Device
摘要
伴隨數位科技化時代演進，智慧型手機已成為多數人日常生活不可或缺之物品，惟在享受便捷之餘，同時也招引新型態詐騙手法的威脅。在2023年8月，國內警方揭露一起利用語音深度偽造假冒超商主管致電內部員工，要求操作機器及購買遊戲點數的詐騙案例，在國外亦於數年前即有類似事件發生，彰顯了智慧型手機用戶既使在面對熟識來電語音，仍潛在「耳聽不能為憑」的風險。面對前揭語音深度偽造詐騙案件，觀諸當前國內相關業者推出的來電辨識功能，多僅透過「來電號碼」辨識是否為詐騙、騷擾電話，完全無法針對來電即時語音進行辨識功能；因此，本研究提出一項預警防制方式，旨在民眾接獲智慧型手機來電當下，同步結合準確度達98.1%的深度學習技術手機APP進行語音特徵掃描，於接通後即時分析該通來電語音遭深偽變造之機率，並透過彈出式警告視窗警示民眾，藉以提高警覺進一步查證，避免落入設計精巧的語音詐騙陷阱。雖然近年政府機關已關注到AI詐騙並向民眾多方宣導，但難以期望均由民眾自主保持警覺性，遂藉由建構本項智慧型手機語音偽造詐騙預警系統提供民眾辨識資訊，有效增進民眾面對新型態電信詐騙的防制契機。
並列摘要
Accompanying the evolution of the digital technology era, smartphones have become indispensable items in most people's daily lives. However, while enjoying the convenience, new types of fraud also pose threats. In August 2023, domestic police uncovered a fraud case involving audio deepfake, where scammers impersonated a convenience store supervisor to call internal employees, instructing them to operate machines and purchase game points. Similar incidents have occurred abroad in recent years, highlighting the risk that smartphone users face even when receiving calls from familiar voices, as "hearing is not believing." In the face of such audio deepfake fraud cases, current domestic call recognition functions primarily identify fraud or harassment calls based on "caller numbers" and cannot recognize real-time call audio. Therefore, this research proposes a precautionary prevention method.When users receive smartphone calls, a mobile app using deep learning technology with an accuracy rate of 98.1% scans the voice features. It instantly analyzes the probability of the call audio being deepfake after connecting, and alerts users with a pop-up warning window to increase vigilance and verify further, avoiding sophisticated audio fraud traps. Although government agencies have paid attention to AI fraud and promoted awareness among the public in recent years, it is unrealistic to expect everyone to remain vigilant autonomously. Thus, by constructing this smartphone deepfake audio fraud alert system, we provide users with recognition information, effectively enhancing their chances to prevent new types of telecom fraud.