扁平化办公白领职场人士通话场景元素

生成一张TASLP论文Introduction用“Motivation figure”，强调动机链而非方法细节。风格：极简矢量扁平、白底、少文字、可黑白打印。布局：2×2四宫格，每格左上角标(a)(b)(c)(d)。(a) Threat/Need：画电话/客服/金融图标+“Deepfake attack”红色警告；文字：Audio deepfakes threaten real applications → need ADD.(b) Data Reality (Centralized infeasible)：画多个机构的“raw audio waveform”指向中央服务器的箭头被红叉阻断；旁边加锁/盾牌；再加一句：Real speech is private & regulated; centralized collection also risks misuse for voice cloning.(c) Naive FL is not enough：画设备→云端FedAvg，仅上传model updates；旁边用散点不均表示Non-IID，再加掉线符号表示unstable participation；文字：Non-IID + dropouts → performance drop / instability (EER↑).(d) Goal (This paper)：画“Privacy-constrained training”框：raw audio留在设备（data stays local），云端输出Robust Deepfake Detector；绿色对勾：High accuracy & stable under privacy constraint. 底部统一一句：Goal: Train a robust deepfake detector without sharing raw audio.