upgrade gradio

.gradio/certificate.pem

@@ -0,0 +1,31 @@
+-----BEGIN CERTIFICATE-----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=
+-----END CERTIFICATE-----

app.py

@@ -28,34 +28,122 @@ yolom = YOLO(MODEL_PATH)  # wrapper
 # yolom_c = YOLO(MODEL_PATH_C)  # wrapper
 # put underlying module to eval on correct device might be needed in attacks functions
   
-def run_detection_on_pil(img_pil: Image.Image, eval_model_state, conf: float = 0.45):
+# def run_detection_on_pil(img_pil: Image.Image, eval_model_state, conf: float = 0.45):
+#     """
+#     Use ultralytics wrapper predict to get a visualization image with boxes.
+#     This is inference-only and does not require gradient.
+#     """
+#     # ultralytics accepts numpy array (H,W,3) in RGB, we pass it directly
+#     img = np.array(img_pil)
+#     # use model.predict with verbose=False to avoid prints
+#     eva_model =  yolom if eval_model_state == "yolom" else YOLO(MODEL_PATH_C)
+#     res = eva_model.predict(source=img, conf=conf, imgsz=imgsz, save=False, verbose=False)
+#     r = res[0]
+#     im_out = img.copy()
+#     # Boxes object may be empty
+#     try:
+#         boxes = r.boxes
+#         for box in boxes:
+#             xyxy = box.xyxy[0].cpu().numpy().astype(int)
+#             x1, y1, x2, y2 = map(int, xyxy)
+#             conf_score = float(box.conf[0].cpu().numpy())
+#             cls_id = int(box.cls[0].cpu().numpy())
+#             # label = f"{cls_id}:{conf_score:.2f}"
+#             label = f"{names[cls_id]}:{conf_score:.2f}"
+#             cv2.rectangle(im_out, (x1, y1), (x2, y2), (0,255,0), 2)
+#             cv2.putText(im_out, label, (x1, max(10,y1-5)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 1)
+#     except Exception as e:
+#         # if no boxes or structure unexpected, just return original
+#         pass
+    # return Image.fromarray(im_out)
+def iou(a, b):
+    ax1, ay1, ax2, ay2 = a
+    bx1, by1, bx2, by2 = b
+    iw = max(0, min(ax2, bx2) - max(ax1, bx1))
+    ih = max(0, min(ay2, by2) - max(ay1, by1))
+    inter = iw * ih
+    if inter <= 0:
+        return 0.0
+    area_a = max(0, ax2 - ax1) * max(0, ay2 - ay1)
+    area_b = max(0, bx2 - bx1) * max(0, by2 - by1)
+    return inter / (area_a + area_b - inter + 1e-9)
+    
+# def center_and_diag(b): #IOU足够好 未启用
+#     x1, y1, x2, y2 = b
+#     cx = 0.5 * (x1 + x2); cy = 0.5 * (y1 + y2)
+#     diag = max(1e-9, ((x2 - x1)**2 + (y2 - y1)**2)**0.5)
+#     area = max(0, (x2 - x1)) * max(0, (y2 - y1))
+#     return cx, cy, diag, area
+  
+def run_detection_on_pil(img_pil: Image.Image, eval_model_state, conf: float = 0.45, GT_boxes=None):
     """
-    Use ultralytics wrapper predict to get a visualization image with boxes.
-    This is inference-only and does not require gradient.
+    推理+可视化。GT_boxes 和返回的 preds 都是:
+      [{'xyxy': (x1,y1,x2,y2), 'cls': int, 'conf': float(optional)}]
     """
-    # ultralytics accepts numpy array (H,W,3) in RGB, we pass it directly
+    import numpy as np, cv2, math
+    from ultralytics import YOLO
+
+    # ---- 1) 推理 ----
     img = np.array(img_pil)
-    # use model.predict with verbose=False to avoid prints
-    eva_model =  yolom if eval_model_state == "yolom" else YOLO(MODEL_PATH_C)
+    eva_model = yolom if eval_model_state == "yolom" else YOLO(MODEL_PATH_C)
     res = eva_model.predict(source=img, conf=conf, imgsz=imgsz, save=False, verbose=False)
     r = res[0]
     im_out = img.copy()
-    # Boxes object may be empty
+
+    # 名称表（尽量稳）
+    names = getattr(r, "names", None)
+    if names is None and hasattr(eva_model, "model") and hasattr(eva_model.model, "names"):
+        names = eva_model.model.names
+
+    # ---- 2) 规整预测框到简单结构 ----
+    preds = []
     try:
-        boxes = r.boxes
-        for box in boxes:
-            xyxy = box.xyxy[0].cpu().numpy().astype(int)
-            x1, y1, x2, y2 = map(int, xyxy)
-            conf_score = float(box.conf[0].cpu().numpy())
-            cls_id = int(box.cls[0].cpu().numpy())
-            # label = f"{cls_id}:{conf_score:.2f}"
-            label = f"{names[cls_id]}:{conf_score:.2f}"
-            cv2.rectangle(im_out, (x1, y1), (x2, y2), (0,255,0), 2)
-            cv2.putText(im_out, label, (x1, max(10,y1-5)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 1)
+        bxs = r.boxes
+        if bxs is not None and len(bxs) > 0:
+            for b in bxs:
+                xyxy = b.xyxy[0].detach().cpu().numpy().tolist()
+                x1, y1, x2, y2 = [int(v) for v in xyxy]
+                cls_id = int(b.cls[0].detach().cpu().numpy())
+                conf_score = float(b.conf[0].detach().cpu().numpy())
+                preds.append({'xyxy': (x1, y1, x2, y2), 'cls': cls_id, 'conf': conf_score})
     except Exception as e:
-        # if no boxes or structure unexpected, just return original
-        pass
-    return Image.fromarray(im_out)
+        print("collect preds error:", e)
+
+    # ---- 3) IoU 匹配 + 画框 ----
+    IOU_THR = 0.3
+    # CENTER_DIST_RATIO = 0.30    # 中心点距离 / 预测框对角线 <= 0.30 即视为同一目标
+    # AREA_RATIO_THR = 0.25       # 面积比例下限：min(area_p, area_g) / max(...) >= 0.25
+    gt_used = set()
+    
+    for p in preds:
+        color = (0, 255, 0)   # 同类：绿
+        px1, py1, px2, py2 = p['xyxy']
+        pname = names[p['cls']] if (names is not None and p['cls'] in getattr(names, 'keys', lambda: [])()) else (
+                names[p['cls']] if (isinstance(names, (list, tuple)) and 0 <= p['cls'] < len(names)) else str(p['cls'])
+        )
+        label = f"{pname}:{p.get('conf', 0.0):.2f}"
+
+        if GT_boxes != None:
+            # 找 IoU 最高的未用 GT
+            best_j, best_iou = -1, 0.0
+            for j, g in enumerate(GT_boxes):
+                if j in gt_used:
+                    continue
+                i = iou(p['xyxy'], g['xyxy'])
+                if i > best_iou:
+                    best_iou, best_j = i, j
+    
+            # 颜色规则：匹配且同类=绿；匹配但异类=红；
+            if best_iou >= IOU_THR:
+                gt_used.add(best_j)
+                if p['cls'] != int(GT_boxes[best_j]['cls']):
+                    color = (255, 0, 0)   # 异类：红
+
+        cv2.rectangle(im_out, (px1, py1), (px2, py2), color, 2)
+        cv2.putText(im_out, label, (px1, max(10, py1 - 5)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 1)
+
+    return Image.fromarray(im_out), preds
+
 
 def detect_and_attack(image, eval_model_state, attack_mode, eps, alpha, iters, conf):
     if image is None:
@@ -63,7 +151,7 @@ def detect_and_attack(image, eval_model_state, attack_mode, eps, alpha, iters, c
 
     pil = Image.fromarray(image.astype('uint8'), 'RGB')
 
-    original_vis = run_detection_on_pil(pil, eval_model_state, conf=conf) 
+    original_vis, GT_boxes = run_detection_on_pil(pil, eval_model_state, conf=conf, GT_boxes=None) 
 
     if attack_mode == "none":
         return original_vis, None
@@ -79,7 +167,7 @@ def detect_and_attack(image, eval_model_state, attack_mode, eps, alpha, iters, c
         print("Whitebox attack failed:", ex)
         adv_pil = attacks.demo_random_perturbation(pil, eps=eps)
 
-    adv_vis = run_detection_on_pil(adv_pil, eval_model_state, conf=conf)
+    adv_vis, _ = run_detection_on_pil(adv_pil, eval_model_state, conf=conf, GT_boxes=GT_boxes)
     return original_vis, adv_vis
 
 
@@ -190,8 +278,14 @@ if __name__ == "__main__":
             outputs=[out_orig, out_adv]
         )
 
-    # demo.queue(concurrency_count=2, max_size=20)
-    demo.launch()
-    # demo.launch(server_name="0.0.0.0", server_port=7860)
+    demo.queue(default_concurrency_limit=2, max_size=20)
+    if os.getenv("SPACE_ID"):
+        demo.launch(
+            server_name="0.0.0.0",
+            server_port=int(os.getenv("PORT", 7860)),
+            show_error=True,
+        )
+    else:
+        demo.launch()
 
     

app_old.py

@@ -0,0 +1,197 @@
+import io
+import numpy as np
+from PIL import Image
+import gradio as gr
+import torch
+from ultralytics import YOLO
+import cv2
+import attacks  # 上面那个 attacks.py，确保和 app.py 在同一目录或可 import 的包路径
+import os, glob
+
+
+# MODEL_PATH = "weights/yolov8s_3cls.pt"
+MODEL_PATH = "weights/fed_model2.pt"
+MODEL_PATH_C = "weights/yolov8s_3cls.pt"
+
+names = ['car', 'van', 'truck']
+imgsz = 640
+
+SAMPLE_DIR = "./images/train"
+SAMPLE_IMAGES = sorted([
+    p for p in glob.glob(os.path.join(SAMPLE_DIR, "*"))
+    if os.path.splitext(p)[1].lower() in [".jpg", ".jpeg", ".png", ".bmp", ".webp"]
+])[:4]  # 只取前4张
+
+# Load ultralytics model (wrapper)
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+yolom = YOLO(MODEL_PATH)  # wrapper
+# yolom_c = YOLO(MODEL_PATH_C)  # wrapper
+# put underlying module to eval on correct device might be needed in attacks functions
+  
+def run_detection_on_pil(img_pil: Image.Image, eval_model_state, conf: float = 0.45):
+    """
+    Use ultralytics wrapper predict to get a visualization image with boxes.
+    This is inference-only and does not require gradient.
+    """
+    # ultralytics accepts numpy array (H,W,3) in RGB, we pass it directly
+    img = np.array(img_pil)
+    # use model.predict with verbose=False to avoid prints
+    eva_model =  yolom if eval_model_state == "yolom" else YOLO(MODEL_PATH_C)
+    res = eva_model.predict(source=img, conf=conf, imgsz=imgsz, save=False, verbose=False)
+    r = res[0]
+    im_out = img.copy()
+    # Boxes object may be empty
+    try:
+        boxes = r.boxes
+        for box in boxes:
+            xyxy = box.xyxy[0].cpu().numpy().astype(int)
+            x1, y1, x2, y2 = map(int, xyxy)
+            conf_score = float(box.conf[0].cpu().numpy())
+            cls_id = int(box.cls[0].cpu().numpy())
+            # label = f"{cls_id}:{conf_score:.2f}"
+            label = f"{names[cls_id]}:{conf_score:.2f}"
+            cv2.rectangle(im_out, (x1, y1), (x2, y2), (0,255,0), 2)
+            cv2.putText(im_out, label, (x1, max(10,y1-5)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 1)
+    except Exception as e:
+        # if no boxes or structure unexpected, just return original
+        pass
+    return Image.fromarray(im_out)
+
+def detect_and_attack(image, eval_model_state, attack_mode, eps, alpha, iters, conf):
+    if image is None:
+        return None, None
+
+    pil = Image.fromarray(image.astype('uint8'), 'RGB')
+
+    original_vis = run_detection_on_pil(pil, eval_model_state, conf=conf) 
+
+    if attack_mode == "none":
+        return original_vis, None
+
+    try:
+        if attack_mode == "fgsm":
+            adv_pil = attacks.fgsm_attack_on_detector(yolom, pil, eps=eps, device=device, imgsz=imgsz)
+        elif attack_mode == "pgd":
+            adv_pil = attacks.pgd_attack_on_detector(yolom, pil, eps=eps, alpha=alpha, iters=iters, device=device, imgsz=imgsz)
+        else:
+            adv_pil = attacks.demo_random_perturbation(pil, eps=eps)
+    except Exception as ex:
+        print("Whitebox attack failed:", ex)
+        adv_pil = attacks.demo_random_perturbation(pil, eps=eps)
+
+    adv_vis = run_detection_on_pil(adv_pil, eval_model_state, conf=conf)
+    return original_vis, adv_vis
+
+
+# Gradio UI
+if __name__ == "__main__":
+    title = "Federated Adversarial Attack — FGSM/PGD Demo"
+    desc_html = (
+        "Adversarial examples are generated locally using a "
+        "<strong>client-side</strong> model’s gradients (white-box), then evaluated against the "
+        "<strong>server-side aggregated (FedAvg) central model</strong>. "
+        "If the perturbation transfers, it can "
+        "degrade or alter the FedAvg model’s predictions on the same input image."
+    )
+    with gr.Blocks(title=title) as demo:
+        # 标题居中
+        gr.Markdown(f"""
+            <div>
+              <h1 style='text-align:center;margin-bottom:0.2rem'>{title}</h1>
+              <p style='opacity:0.85'>{desc_html}</p>
+            </div>""")
+
+        with gr.Row():
+            # ===== 左列：两个输入区块 =====
+            with gr.Column(scale=5):
+                # 输入区块 1：上传窗口 & 样例选择 —— 左右并列
+                with gr.Row():
+                    with gr.Column(scale=7):
+                        in_img = gr.Image(type="numpy", label="Input image")
+                    with gr.Column(scale=2):
+                        if SAMPLE_IMAGES:
+                            gr.Examples(
+                                examples=SAMPLE_IMAGES,
+                                inputs=[in_img],
+                                label=f"Select from sample images",
+                                examples_per_page=4,
+                                # run_on_click 默认为 False（只填充，不执行）
+                            )
+
+                # 输入 2：攻击与参数
+                with gr.Accordion("Attack mode", open=True):
+                    attack_mode = gr.Radio(
+                        choices=["none", "fgsm", "pgd", "random noise"],
+                        value="fgsm",
+                        label="",          
+                        show_label=False
+                    )
+                    eps   = gr.Slider(0.0, 0.3, step=0.01, value=0.0314, label="eps")
+                    alpha = gr.Slider(0.001, 0.05, step=0.001, value=0.0078, label="alpha (PGD step)")
+                    iters = gr.Slider(1, 100, step=1, value=10, label="PGD iterations")
+                    conf  = gr.Slider(0.0, 1.0, step=0.01, value=0.45, label="Confidence threshold (live)")
+
+                with gr.Row():
+                    btn_clear  = gr.ClearButton(
+                        components=[in_img, eps, alpha, iters, conf],  # 不清空 attack_mode
+                        value="Clear"
+                    )
+                    btn_submit = gr.Button("Submit", variant="primary")
+
+            # ===== 右列：两个输出区块 =====
+            with gr.Column(scale=5):
+                # 新增：评测模型选择
+                with gr.Row():
+                    eval_choice = gr.Dropdown(
+                        choices=[(f"Client model {MODEL_PATH}", "client"),
+                                 (f"Central model {MODEL_PATH_C}", "central")],
+                        value="client",                 # ★ 初始值为合法 value
+                        label="Evaluation model"
+                    )
+                    
+                    eval_model_state = gr.State(value="yolom")  
+
+                # ★ 合并后的单一回调：规范化下拉值 + 返回(更新后的下拉值, 模型对象)
+                def on_eval_change(val: str):
+                    if isinstance(val, (list, tuple)):
+                        val = val[0] if len(val) else "client"
+                    if val not in ("client", "central"):
+                        val = "client"
+                    model = "yolom" if val == "client" else "yolom_c"
+                    return gr.update(value=val), model
+
+                # 页面加载时同步一次，避免初次为空/不一致
+                demo.load(
+                    fn=on_eval_change,
+                    inputs=eval_choice,
+                    outputs=[eval_choice, eval_model_state]
+                )
+
+                # 仅这一条 change 绑定（删掉你原来那个只写 State 的 change，避免并发覆盖）
+                eval_choice.change(
+                    fn=on_eval_change,
+                    inputs=eval_choice,
+                    outputs=[eval_choice, eval_model_state]
+                )
+                out_orig = gr.Image(label="Original detection")
+                out_adv  = gr.Image(label="After attack detection")
+
+        # Submit：手动运行
+        btn_submit.click(
+            fn=detect_and_attack,
+            inputs=[in_img, eval_model_state, attack_mode, eps, alpha, iters, conf],
+            outputs=[out_orig, out_adv]
+        )
+
+        # 仅 conf 滑块“实时”
+        conf.release(
+            fn=detect_and_attack,
+            inputs=[in_img, eval_model_state, attack_mode, eps, alpha, iters, conf],
+            outputs=[out_orig, out_adv]
+        )
+
+    # demo.queue(concurrency_count=2, max_size=20)
+    demo.launch()
+    # demo.launch(server_name="0.0.0.0", server_port=7860)
+
+    

requirements.txt

@@ -1,7 +1,7 @@
 # --- Core ---
-gradio==4.44.1
-gradio_client==0.16.6
-huggingface_hub==0.25.2     
+gradio==5.49.1
+gradio_client>=1.13.0,<2
+huggingface_hub>=0.29.0
 ultralytics>=8.3.0
 torch>=2.2.0
 torchvision>=0.17.0