what.examples.yolov3_pcb_attack_demo

  1import cv2
  2import numpy as np
  3
  4from what.models.detection.datasets.coco import COCO_CLASS_NAMES
  5from what.models.detection.utils.box_utils import draw_bounding_boxes
  6from what.models.detection.yolo.utils.yolo_utils import yolo_process_output, yolov3_anchors, yolov3_tiny_anchors
  7
  8from what.attacks.detection.yolo.PCB import PCBAttack
  9from what.utils.resize import bilinear_resize
 10
 11from what.cli.model import *
 12from what.utils.file import get_file
 13
 14def yolov3_pcb_attack_demo():
 15
 16    # Target Model
 17    what_yolov3_model_list = what_model_list[0:4]
 18
 19    classes = COCO_CLASS_NAMES
 20
 21    colors = np.random.uniform(0, 255, size=(len(classes), 3))
 22
 23    # Check what_model_list for all supported models
 24    index = 3
 25
 26    # Download the model first if not exists
 27    if not os.path.isfile(os.path.join(WHAT_MODEL_PATH, what_yolov3_model_list[index][WHAT_MODEL_FILE_INDEX])):
 28        get_file(what_yolov3_model_list[index][WHAT_MODEL_FILE_INDEX],
 29                    WHAT_MODEL_PATH,
 30                    what_yolov3_model_list[index][WHAT_MODEL_URL_INDEX],
 31                    what_yolov3_model_list[index][WHAT_MODEL_HASH_INDEX])
 32
 33    # Adversarial Attack
 34    model_path = os.path.join(WHAT_MODEL_PATH, what_yolov3_model_list[index][WHAT_MODEL_FILE_INDEX])
 35    attack = PCBAttack(model_path, "multi_untargeted", classes, decay=0.99)
 36    attack.fixed = False
 37
 38    last_outs = None
 39    last_boxes = None
 40    last_probs = None
 41
 42    # Initialize the camera
 43    video = input(f"Please input the OpenCV capture device (e.g. 0, 1, 2): ")
 44
 45    while not video.isdigit():
 46        video = input(f"Please input the OpenCV capture device (e.g. 0, 1, 2): ")
 47
 48    # Capture from camera
 49    cap = cv2.VideoCapture(int(video))
 50    #cap.set(3, 1920)
 51    #cap.set(4, 1080)
 52
 53    while(True): 
 54        # Capture the video frame
 55        success, origin_cv_image = cap.read()  # read the camera frame
 56        if not success:
 57            break
 58
 59        # For YOLO, the input pixel values are normalized to [0, 1]
 60        input_cv_image = cv2.resize(origin_cv_image, (416, 416))
 61        input_cv_image = np.array(input_cv_image).astype(np.float32) / 255.0
 62
 63        # Image preprocessing
 64        input_cv_image = cv2.cvtColor(input_cv_image, cv2.COLOR_BGR2RGB)
 65
 66        # Yolo inference
 67        input_cv_image, outs = attack.attack(input_cv_image)
 68
 69        if last_outs is not None:
 70            res_list = []
 71            for out, last_out in zip(outs, last_outs):
 72                out = out.reshape((-1, 5 + len(classes)))
 73                last_out = last_out.reshape((-1, 5 + len(classes)))
 74
 75                res = np.mean(out[:, 4] - last_out[:, 4])
 76                res_list.append(res)
 77
 78        else:
 79            last_outs = outs
 80
 81        boxes, labels, probs = yolo_process_output(outs, yolov3_tiny_anchors, len(classes))
 82
 83        if last_boxes is not None:
 84            # Eliminate the boxes with low confidence and overlaped boxes
 85            if last_boxes.size > 0 and boxes.size > 0:
 86                indexes = cv2.dnn.NMSBoxes(np.vstack((boxes, last_boxes)).tolist(), np.hstack((np.array(probs), np.array(last_probs))), 0.5, 0.4)
 87                if len(indexes) > 0:
 88                    indexes = indexes.flatten()
 89
 90        last_boxes = np.copy(boxes)
 91        last_probs = np.copy(probs)
 92
 93        # Draw bounding boxes
 94        out_img = cv2.cvtColor(origin_cv_image, cv2.COLOR_BGR2RGB)
 95        out_img = out_img.astype(np.float32) / 255.0
 96        height, width, _ = out_img.shape
 97        noise = attack.noise
 98
 99        # Resize the noise to the same shape as the input image
100        # noise_r = bilinear_resize(noise[:, :, 0], height, width)
101        # noise_g = bilinear_resize(noise[:, :, 1], height, width)
102        # noise_b = bilinear_resize(noise[:, :, 2], height, width)
103        # noise = np.dstack((noise_r, noise_g, noise_b))
104
105        # out_img = out_img + noise
106        out_img = np.clip(out_img, 0, 1)
107
108        out_img = (out_img * 255.0).astype(np.uint8)
109
110        # for i in range(boxes.shape[0]):
111        #     logger.info(f"{classes[labels[i]]}: {probs[i]:.2f}")
112
113        out_img = cv2.cvtColor(out_img, cv2.COLOR_RGB2BGR)
114        if len(boxes) > 0:
115            out_img = draw_bounding_boxes(out_img, boxes, labels, classes, probs);
116
117        cv2.namedWindow("result", cv2.WINDOW_NORMAL)
118        cv2.imshow("result", out_img)
119
120        if (cv2.waitKey(1) & 0xFF == ord('q')):
121            break