3131
3232 < div class ="nav-links ">
3333 < a href ="#features "> Features</ a >
34+ < a href ="#partial-deepfake "> Partial Deepfake</ a >
3435 < div class ="dropdown ">
3536 < a href ="#modules " class ="dropdown-trigger "> Modules < i class ="fas fa-chevron-down "> </ i > </ a >
3637 < div class ="dropdown-content ">
6465 < div class ="hero-content mx-auto ">
6566 < div class ="badge badge-white "> < i class ="fas fa-fire "> </ i > v0.2 — 455+ Pretrained Models on HuggingFace</ div >
6667 < h1 class ="text-white "> Build, Train & Evaluate< br > Deepfake Audio Detectors</ h1 >
67- < p class ="text-white-opacity "> A modular, configuration-driven framework. Mix frontends, backends, and losses via YAML — no code changes needed .</ p >
68+ < p class ="text-white-opacity "> Clip-level detection and partial deepfake localization — one YAML config, swap frontends, backends, and losses with no code changes.</ p >
6869 < div class ="hero-buttons justify-center ">
6970 < a href ="#quickstart " class ="btn btn-white btn-lg "> Get Started</ a >
7071 < a href ="https://huggingface.co/DeepFense " target ="_blank " class ="btn btn-outline-hero btn-lg ">
@@ -107,6 +108,57 @@ <h1 class="text-white">Build, Train & Evaluate<br>Deepfake Audio Detectors</
107108 </ div >
108109 </ section >
109110
111+ <!-- Partial Deepfake Section -->
112+ < section id ="partial-deepfake " class ="partial-section ">
113+ < div class ="container ">
114+ < div class ="section-header ">
115+ < div class ="badge badge-partial "> < i class ="fas fa-location-dot "> </ i > Partial deepfake localization</ div >
116+ < h2 > Two detection modes</ h2 >
117+ < p > Same modular pipeline — different handling of the time axis.</ p >
118+ </ div >
119+
120+ < div class ="modes-grid ">
121+ < div class ="mode-card ">
122+ < span class ="mode-tag "> Default branch: < code > master</ code > </ span >
123+ < h3 > Clip-level</ h3 >
124+ < p > One label per utterance. Is the whole clip real or fake?</ p >
125+ < ul class ="mode-flow ">
126+ < li > Raw audio → Frontend → Backend < strong > pools time</ strong > </ li >
127+ < li > CrossEntropy / OC-Softmax → one score</ li >
128+ < li > Metrics: EER, minDCF, ACC</ li >
129+ </ ul >
130+ </ div >
131+
132+ < div class ="mode-card mode-card-highlight ">
133+ < span class ="mode-tag "> Branch: < code > deepfense-partial</ code > </ span >
134+ < h3 > Partial deepfake</ h3 >
135+ < p > Dense frame labels. < strong > Where</ strong > is the spoof in the clip?</ p >
136+ < ul class ="mode-flow ">
137+ < li > Raw audio → Frontend → FrameMLP / GMLP < strong > keeps time</ strong > </ li >
138+ < li > FramewiseCrossEntropy → score per frame</ li >
139+ < li > Metrics: FRAME_F1, < strong > Range EER</ strong > , Segment EER, MULTIRES_EER</ li >
140+ </ ul >
141+ </ div >
142+ </ div >
143+
144+ < div class ="partial-cta ">
145+ < p > For PartialSpoof-style per-frame training and localization metrics, switch to the partial branch:</ p >
146+ < div class ="code-block partial-code-block ">
147+ < div class ="code-header "> Partial deepfake setup</ div >
148+ < pre > < code class ="bash "> git clone https://github.com/Yaselley/deepfense-framework
149+ cd deepfense-framework
150+ git fetch origin
151+ git checkout deepfense-partial
152+ pip install -e .
153+
154+ deepfense train --config deepfense/config/experiments/temporal_deepfake_example.yaml</ code > </ pre >
155+ < button class ="copy-btn "> < i class ="far fa-copy "> </ i > </ button >
156+ </ div >
157+ < p class ="partial-note "> < code > master</ code > covers clip-level detection only. Partial deepfake configs, < code > TemporalDetector</ code > , and Range EER live on < code > deepfense-partial</ code > .</ p >
158+ </ div >
159+ </ div >
160+ </ section >
161+
110162 <!-- Features Section -->
111163 < section id ="features " class ="features ">
112164 < div class ="container ">
@@ -145,7 +197,7 @@ <h3>HuggingFace Integration</h3>
145197 < i class ="fas fa-chart-bar "> </ i >
146198 </ div >
147199 < h3 > Standardized Metrics</ h3 >
148- < p > Automatic tracking of EER, minDCF, actDCF, F1, and Accuracy with WandB logging and checkpointing .</ p >
200+ < p > Clip-level: EER, minDCF, actDCF, F1. Partial branch: framewise metrics plus Range EER and Segment EER for localization .</ p >
149201 </ div >
150202
151203 < div class ="card ">
@@ -200,6 +252,8 @@ <h3>Backends</h3>
200252 < li > RawNet2</ li >
201253 < li > MLP</ li >
202254 < li > TCM</ li >
255+ < li > FrameMLP</ li >
256+ < li > GMLP</ li >
203257 </ ul >
204258 </ div >
205259
@@ -212,6 +266,7 @@ <h3>Loss Functions</h3>
212266 < li > OC-Softmax</ li >
213267 < li > AM-Softmax</ li >
214268 < li > A-Softmax</ li >
269+ < li > FramewiseCE</ li >
215270 </ ul >
216271 </ div >
217272
@@ -238,9 +293,9 @@ <h3>Metrics</h3>
238293 < ul class ="module-list ">
239294 < li > EER</ li >
240295 < li > minDCF</ li >
241- < li > actDCF </ li >
242- < li > ACC </ li >
243- < li > F1 </ li >
296+ < li > FRAME_F1 </ li >
297+ < li > Range EER </ li >
298+ < li > Segment EER </ li >
244299 </ ul >
245300 </ div >
246301
0 commit comments