| import spaces |
| import gradio as gr |
| import torch |
| import tempfile |
| import os |
| from PIL import Image |
| from transformers import pipeline |
|
|
| |
| MODELS = { |
| ("fast", "kodf"): "KoreaPeter/ms-eff-gcvit-deepfake-b0-kodf", |
| ("fast", "celeb-df-v2"): "KoreaPeter/ms-eff-gcvit-deepfake-b0-celeb-df-v2", |
| ("fast", "ff++"): "KoreaPeter/ms-eff-gcvit-deepfake-b0-ff-plus-plus", |
| ("pro", "kodf"): "KoreaPeter/ms-eff-gcvit-deepfake-b5-kodf", |
| ("pro", "celeb-df-v2"): "KoreaPeter/ms-eff-gcvit-deepfake-b5-celeb-df-v2", |
| ("pro", "ff++"): "KoreaPeter/ms-eff-gcvit-deepfake-b5-ff-plus-plus", |
| } |
|
|
| _cache: dict = {} |
|
|
| def _get_pipeline(variant: str, dataset: str): |
| key = (variant, dataset) |
| if key not in _cache: |
| device = "cuda" if torch.cuda.is_available() else "cpu" |
| _cache[key] = pipeline( |
| "image-classification", |
| model=MODELS[key], |
| trust_remote_code=True, |
| device=device, |
| ) |
| return _cache[key] |
|
|
|
|
| |
| @spaces.GPU |
| def predict( |
| image: Image.Image, |
| variant: str, |
| dataset: str, |
| margin_ratio: float, |
| conf_thres: float, |
| min_face_ratio: float, |
| tta_hflip: float, |
| top_k: int, |
| ): |
| if image is None: |
| return None, "⚠️ Please upload an image." |
|
|
| with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f: |
| image.save(f.name) |
| tmp_path = f.name |
|
|
| try: |
| pipe = _get_pipeline(variant, dataset) |
| results = pipe( |
| tmp_path, |
| margin_ratio=margin_ratio, |
| conf_thres=conf_thres, |
| min_face_ratio=min_face_ratio, |
| tta_hflip=tta_hflip, |
| top_k=top_k, |
| ) |
| except Exception as e: |
| return None, f"❌ Pipeline error: {e}" |
| finally: |
| os.unlink(tmp_path) |
|
|
| if not results: |
| return None, "⚠️ No result returned." |
|
|
| top = results[0] |
| if top.get("label") == "no_face_detected": |
| return ( |
| {"REAL": 0.0, "FAKE": 0.0}, |
| "😶 No face detected.\nPlease use an image with a clearly visible face.", |
| ) |
|
|
| scores = {r["label"].upper(): round(r["score"], 4) for r in results |
| if r.get("label") in ("fake", "real")} |
|
|
| fake_prob = scores.get("FAKE", 0.0) |
|
|
| if fake_prob >= 0.7: |
| verdict = f"🔴 FAKE ({fake_prob * 100:.1f}% confidence)" |
| elif fake_prob >= 0.4: |
| verdict = f"🟡 UNCERTAIN (FAKE {fake_prob * 100:.1f}%)" |
| else: |
| verdict = f"🟢 REAL ({(1 - fake_prob) * 100:.1f}% confidence)" |
|
|
| return scores, verdict |
|
|
|
|
| |
| HERO = """ |
| <div style=" |
| background: linear-gradient(135deg, #c2185b 0%, #e91e63 45%, #880e4f 100%); |
| border-radius: 12px; padding: 36px 40px; margin-bottom: 8px; |
| border: 1px solid #ad1457; box-shadow: 0 4px 24px rgba(194,24,91,0.4); |
| "> |
| <div style="display:flex; align-items:center; gap:16px; margin-bottom:12px;"> |
| <span style="font-size:2.4rem;">🕵️</span> |
| <div> |
| <h1 style="margin:0; color:#fce4ec; font-size:1.9rem; font-weight:700; letter-spacing:-0.5px;">DeepGuard</h1> |
| <p style="margin:4px 0 0; color:#f48fb1; font-size:1rem;">Image Deepfake Detection</p> |
| </div> |
| </div> |
| <p style="margin:0 0 16px; color:#f8bbd0; font-size:0.95rem; line-height:1.6;"> |
| Deepfake detection powered by <strong style="color:#fce4ec;">MS-EffGCViT</strong> |
| · 6 domain-specific models · Asian · Celebrity · Western faces |
| </p> |
| <div style="display:flex; gap:12px; flex-wrap:wrap;"> |
| <a href="https://github.com/HanMoonSub/DeepGuard" target="_blank" |
| style="background:#880e4f; color:#f48fb1; padding:7px 16px; border-radius:6px; |
| text-decoration:none; font-size:0.85rem; border:1px solid #e91e63;">🔗 GitHub</a> |
| <a href="https://huggingface.co/KoreaPeter" target="_blank" |
| style="background:#880e4f; color:#f48fb1; padding:7px 16px; border-radius:6px; |
| text-decoration:none; font-size:0.85rem; border:1px solid #e91e63;">🤗 HuggingFace</a> |
| <a href="https://huggingface.co/spaces/KoreaPeter/DeepFake-Video-Detection" target="_blank" |
| style="background:#880e4f; color:#f48fb1; padding:7px 16px; border-radius:6px; |
| text-decoration:none; font-size:0.85rem; border:1px solid #e91e63;">🎬 Video Detection</a> |
| <a href="https://huggingface.co/spaces/KoreaPeter/DeepFake-Detection-XAI" target="_blank" |
| style="background:#880e4f; color:#f48fb1; padding:7px 16px; border-radius:6px; |
| text-decoration:none; font-size:0.85rem; border:1px solid #e91e63;">🔬 DeepFake XAI</a> |
| </div> |
| </div> |
| """ |
|
|
| |
| THEME = gr.themes.Base( |
| primary_hue="pink", |
| neutral_hue="slate", |
| font=gr.themes.GoogleFont("Inter"), |
| ).set( |
| body_background_fill="#1a0010", |
| body_text_color="#fce4ec", |
| block_background_fill="#2d0020", |
| block_border_color="#ad1457", |
| input_background_fill="#1a0010", |
| button_primary_background_fill="#c2185b", |
| button_primary_text_color="#fce4ec", |
| button_primary_border_color="#e91e63", |
| button_primary_background_fill_hover="#e91e63", |
| button_primary_text_color_hover="#ffffff", |
| ) |
|
|
| CSS = """ |
| .result-box textarea { font-size:1.1rem !important; font-weight:bold !important; text-align:center; } |
| footer { display:none !important; } |
| """ |
|
|
| |
| with gr.Blocks(title="DeepGuard Image") as demo: |
|
|
| gr.HTML(HERO) |
|
|
| with gr.Row(): |
| with gr.Column(scale=1, min_width=300): |
| image_input = gr.Image(type="pil", label="Upload Image") |
| with gr.Accordion("🖼️ Example Images — (click to expand)", open=False): |
| gr.Examples( |
| examples=[ |
| "example1.JPG", |
| "example2.JPG", |
| "example3.JPG", |
| "example4.JPG", |
| ], |
| inputs=image_input, |
| label="All samples are FAKE", |
| ) |
|
|
| gr.Markdown("### ⚡ Model Variant") |
| variant_radio = gr.Radio( |
| choices=["fast", "pro"], |
| value="fast", |
| label=None, |
| info="Fast (b0): 8.7M params | Pro (b5): 50.3M params", |
| ) |
|
|
| gr.Markdown("### 🌍 Training Domain") |
| dataset_radio = gr.Radio( |
| choices=[ |
| ("🇰🇷 KoDF — East Asian faces", "kodf"), |
| ("🌟 Celeb-DF-v2 — Celebrity faces", "celeb-df-v2"), |
| ("🌍 FF++ — Western faces", "ff++"), |
| ], |
| value="ff++", |
| label=None, |
| ) |
|
|
| run_btn = gr.Button("🔍 Analyze", variant="primary", size="lg") |
|
|
| with gr.Column(scale=1, min_width=300): |
| verdict_output = gr.Textbox( |
| label="Result", |
| interactive=False, |
| lines=2, |
| elem_classes=["result-box"], |
| ) |
| score_output = gr.Label(label="Class Probabilities", num_top_classes=2) |
|
|
| gr.Markdown(""" |
| --- |
| **Model Guide** |
| |
| | | ⚡ Fast (b0) | 🔥 Pro (b5) | |
| |---|---|---| |
| | **Params** | 8.7M | 50.3M | |
| | **Resolution** | 224×224 | 384×384 | |
| | **Speed** | Faster | Slower | |
| | **Accuracy** | High | Higher | |
| |
| **Domain Guide** |
| |
| | Domain | Training Data | Best For | |
| |---|---|---| |
| | KoDF | 62K real / 175K fake (Korean) | East Asian faces | |
| | Celeb-DF-v2 | 890 real / 5.6K fake (Celebrities) | Celebrity faces | |
| | FF++ | 1K real / 6K fake (4 manipulation types) | Western faces | |
| """) |
|
|
| with gr.Accordion("⚙️ Advanced Settings", open=True): |
| gr.Markdown("Adjust inference parameters. Defaults are recommended for most cases.") |
| with gr.Row(): |
| with gr.Column(): |
| margin_ratio_slider = gr.Slider( |
| minimum=0.0, maximum=0.5, value=0.2, step=0.05, |
| label="Margin Ratio", |
| info="Margin around the detected face bounding box", |
| ) |
| conf_thres_slider = gr.Slider( |
| minimum=0.1, maximum=0.9, value=0.5, step=0.05, |
| label="Face Detection Confidence", |
| info="YOLO face detection confidence threshold", |
| ) |
| with gr.Column(): |
| min_face_ratio_slider = gr.Slider( |
| minimum=0.001, maximum=0.1, value=0.01, step=0.001, |
| label="Min Face Ratio", |
| info="Minimum face-to-frame area ratio to process", |
| ) |
| tta_hflip_slider = gr.Slider( |
| minimum=0.0, maximum=1.0, value=0.0, step=0.1, |
| label="TTA Horizontal Flip", |
| info="⚠️ > 0.0 doubles inference time", |
| ) |
| top_k_slider = gr.Slider( |
| minimum=1, maximum=2, value=2, step=1, |
| label="Top K", |
| info="Number of top predictions to return", |
| ) |
|
|
| run_btn.click( |
| fn=predict, |
| inputs=[ |
| image_input, variant_radio, dataset_radio, |
| margin_ratio_slider, conf_thres_slider, |
| min_face_ratio_slider, tta_hflip_slider, top_k_slider, |
| ], |
| outputs=[score_output, verdict_output], |
| ) |
| image_input.change( |
| fn=lambda img, v, d, mr, ct, mfr, tta, tk: ( |
| predict(img, v, d, mr, ct, mfr, tta, tk) if img is not None else (None, "") |
| ), |
| inputs=[ |
| image_input, variant_radio, dataset_radio, |
| margin_ratio_slider, conf_thres_slider, |
| min_face_ratio_slider, tta_hflip_slider, top_k_slider, |
| ], |
| outputs=[score_output, verdict_output], |
| ) |
|
|
| if __name__ == "__main__": |
| demo.launch(theme=THEME, css=CSS) |
