src/visualization.py

"""
視覺化工具模組

提供對抗樣本視覺化、實驗結果繪圖等功能
"""

import matplotlib.pyplot as plt
import matplotlib
matplotlib.use('Agg')  # 使用非互動式後端
import numpy as np
import torch
from pathlib import Path


def visualize_attack(original, adversarial, perturbation,
                    original_label, orig_pred, adv_pred,
                    epsilon, save_path=None, amplify=10):
    """
    視覺化對抗攻擊結果 (三圖並排)

    Args:
        original (torch.Tensor or np.ndarray): 原始影像
        adversarial (torch.Tensor or np.ndarray): 對抗樣本
        perturbation (torch.Tensor or np.ndarray): 擾動
        original_label (int): 真實標籤
        orig_pred (int): 原始預測
        adv_pred (int): 對抗樣本預測
        epsilon (float): 擾動幅度
        save_path (str, optional): 儲存路徑
        amplify (int): 擾動放大倍數 (方便觀察)
    """
    # 轉換為 numpy
    if torch.is_tensor(original):
        original = original.detach().cpu().numpy()
    if torch.is_tensor(adversarial):
        adversarial = adversarial.detach().cpu().numpy()
    if torch.is_tensor(perturbation):
        perturbation = perturbation.detach().cpu().numpy()

    # 移除 channel 維度
    if original.ndim == 3:
        original = original.squeeze()
    if adversarial.ndim == 3:
        adversarial = adversarial.squeeze()
    if perturbation.ndim == 3:
        perturbation = perturbation.squeeze()

    # 建立圖表
    fig, axes = plt.subplots(1, 3, figsize=(12, 4))

    # 原始影像
    axes[0].imshow(original, cmap='gray')
    axes[0].set_title(f'原始影像\n真實標籤: {original_label}\n預測: {orig_pred}', fontsize=10)
    axes[0].axis('off')

    # 擾動 (放大顯示)
    amplified_pert = perturbation * amplify
    axes[1].imshow(amplified_pert, cmap='seismic', vmin=-epsilon*amplify, vmax=epsilon*amplify)
    axes[1].set_title(f'擾動 (放大 {amplify}x)\nε = {epsilon}', fontsize=10)
    axes[1].axis('off')

    # 對抗樣本
    axes[2].imshow(adversarial, cmap='gray')
    success = "✓ 成功" if orig_pred != adv_pred else "✗ 失敗"
    axes[2].set_title(f'對抗樣本\n預測: {adv_pred}\n攻擊{success}', fontsize=10)
    axes[2].axis('off')

    plt.tight_layout()

    # 儲存或顯示
    if save_path:
        Path(save_path).parent.mkdir(parents=True, exist_ok=True)
        plt.savefig(save_path, dpi=150, bbox_inches='tight')
        print(f"圖片已儲存至 {save_path}")

    plt.close()

    return fig


def visualize_multiple_attacks(examples, epsilon, save_path=None, max_examples=10):
    """
    視覺化多個對抗攻擊範例

    Args:
        examples (list): 攻擊範例列表,每個元素包含:
            - original: 原始影像
            - adversarial: 對抗樣本
            - perturbation: 擾動
            - true_label: 真實標籤
            - original_pred: 原始預測
            - adv_pred: 對抗預測
        epsilon (float): 擾動幅度
        save_path (str, optional): 儲存路徑
        max_examples (int): 最多顯示幾個範例
    """
    n_examples = min(len(examples), max_examples)

    fig, axes = plt.subplots(n_examples, 3, figsize=(10, 3*n_examples))

    if n_examples == 1:
        axes = axes.reshape(1, -1)

    for i, example in enumerate(examples[:n_examples]):
        # 轉換為 numpy
        original = example['original'].squeeze().numpy()
        adversarial = example['adversarial'].squeeze().numpy()
        perturbation = example['perturbation'].squeeze().numpy()

        # 原始影像
        axes[i, 0].imshow(original, cmap='gray')
        if i == 0:
            axes[i, 0].set_title('原始影像', fontsize=12, fontweight='bold')
        axes[i, 0].set_ylabel(f'範例 {i+1}\n真實: {example["true_label"]}\n預測: {example["original_pred"]}',
                             fontsize=10)
        axes[i, 0].set_xticks([])
        axes[i, 0].set_yticks([])

        # 擾動
        axes[i, 1].imshow(perturbation * 10, cmap='seismic', vmin=-epsilon*10, vmax=epsilon*10)
        if i == 0:
            axes[i, 1].set_title(f'擾動 (放大 10x)', fontsize=12, fontweight='bold')
        axes[i, 1].set_xticks([])
        axes[i, 1].set_yticks([])

        # 對抗樣本
        axes[i, 2].imshow(adversarial, cmap='gray')
        if i == 0:
            axes[i, 2].set_title('對抗樣本', fontsize=12, fontweight='bold')
        axes[i, 2].set_ylabel(f'預測: {example["adv_pred"]}', fontsize=10)
        axes[i, 2].set_xticks([])
        axes[i, 2].set_yticks([])

    plt.suptitle(f'FGSM 對抗攻擊範例 (ε = {epsilon})', fontsize=14, fontweight='bold', y=0.995)
    plt.tight_layout()

    if save_path:
        Path(save_path).parent.mkdir(parents=True, exist_ok=True)
        plt.savefig(save_path, dpi=150, bbox_inches='tight')
        print(f"圖片已儲存至 {save_path}")

    plt.close()

    return fig


def plot_accuracy_vs_epsilon(epsilons, accuracies, save_path=None):
    """
    繪製準確率 vs ε 曲線圖

    Args:
        epsilons (list): ε 值列表
        accuracies (list): 對應的準確率列表
        save_path (str, optional): 儲存路徑
    """
    plt.figure(figsize=(10, 6))

    plt.plot(epsilons, accuracies, 'o-', linewidth=2, markersize=8, color='#2E86AB')
    plt.xlabel('擾動幅度 (ε)', fontsize=12, fontweight='bold')
    plt.ylabel('準確率 (%)', fontsize=12, fontweight='bold')
    plt.title('FGSM 攻擊: 準確率 vs 擾動幅度', fontsize=14, fontweight='bold')
    plt.grid(True, alpha=0.3, linestyle='--')

    # 標註數值
    for eps, acc in zip(epsilons, accuracies):
        plt.annotate(f'{acc:.1f}%',
                    xy=(eps, acc),
                    xytext=(0, 10),
                    textcoords='offset points',
                    ha='center',
                    fontsize=9)

    plt.tight_layout()

    if save_path:
        Path(save_path).parent.mkdir(parents=True, exist_ok=True)
        plt.savefig(save_path, dpi=150, bbox_inches='tight')
        print(f"圖片已儲存至 {save_path}")

    plt.close()


def plot_attack_success_rate(epsilons, success_rates, save_path=None):
    """
    繪製攻擊成功率 vs ε 曲線圖

    Args:
        epsilons (list): ε 值列表
        success_rates (list): 攻擊成功率列表
        save_path (str, optional): 儲存路徑
    """
    plt.figure(figsize=(10, 6))

    plt.plot(epsilons, success_rates, 'o-', linewidth=2, markersize=8, color='#A23B72')
    plt.xlabel('擾動幅度 (ε)', fontsize=12, fontweight='bold')
    plt.ylabel('攻擊成功率 (%)', fontsize=12, fontweight='bold')
    plt.title('FGSM 攻擊成功率 vs 擾動幅度', fontsize=14, fontweight='bold')
    plt.grid(True, alpha=0.3, linestyle='--')

    # 標註數值
    for eps, rate in zip(epsilons, success_rates):
        plt.annotate(f'{rate:.1f}%',
                    xy=(eps, rate),
                    xytext=(0, 10),
                    textcoords='offset points',
                    ha='center',
                    fontsize=9)

    plt.tight_layout()

    if save_path:
        Path(save_path).parent.mkdir(parents=True, exist_ok=True)
        plt.savefig(save_path, dpi=150, bbox_inches='tight')
        print(f"圖片已儲存至 {save_path}")

    plt.close()


def plot_comparison(epsilons, original_accs, adversarial_accs, save_path=None):
    """
    繪製原始準確率和對抗準確率的比較圖

    Args:
        epsilons (list): ε 值列表
        original_accs (list): 原始準確率列表
        adversarial_accs (list): 對抗樣本準確率列表
        save_path (str, optional): 儲存路徑
    """
    plt.figure(figsize=(12, 6))

    plt.plot(epsilons, original_accs, 'o-', linewidth=2, markersize=8,
             label='原始準確率', color='#2E86AB')
    plt.plot(epsilons, adversarial_accs, 's-', linewidth=2, markersize=8,
             label='對抗樣本準確率', color='#A23B72')

    plt.xlabel('擾動幅度 (ε)', fontsize=12, fontweight='bold')
    plt.ylabel('準確率 (%)', fontsize=12, fontweight='bold')
    plt.title('FGSM 攻擊效果分析', fontsize=14, fontweight='bold')
    plt.legend(fontsize=11, loc='best')
    plt.grid(True, alpha=0.3, linestyle='--')

    plt.tight_layout()

    if save_path:
        Path(save_path).parent.mkdir(parents=True, exist_ok=True)
        plt.savefig(save_path, dpi=150, bbox_inches='tight')
        print(f"圖片已儲存至 {save_path}")

    plt.close()


def create_gradio_output(original, adversarial, perturbation,
                        original_label, orig_pred, adv_pred, epsilon):
    """
    為 Gradio 介面建立視覺化輸出

    Args:
        original (torch.Tensor): 原始影像
        adversarial (torch.Tensor): 對抗樣本
        perturbation (torch.Tensor): 擾動
        original_label (int): 真實標籤
        orig_pred (int): 原始預測
        adv_pred (int): 對抗預測
        epsilon (float): 擾動幅度

    Returns:
        tuple: (original_dict, adv_dict, visualization_image)
    """
    # 預測結果字典
    original_dict = {f"數字 {orig_pred}": 1.0}
    adv_dict = {f"數字 {adv_pred}": 1.0}

    # 建立視覺化圖片
    fig = visualize_attack(
        original, adversarial, perturbation,
        original_label, orig_pred, adv_pred, epsilon
    )

    # 轉換為 numpy array
    fig.canvas.draw()
    vis_image = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
    vis_image = vis_image.reshape(fig.canvas.get_width_height()[::-1] + (3,))

    plt.close(fig)

    return original_dict, adv_dict, vis_image


if __name__ == "__main__":
    # 測試視覺化功能
    print("測試視覺化模組...")

    # 建立測試資料
    original = torch.rand(1, 28, 28)
    adversarial = torch.clamp(original + torch.randn(1, 28, 28) * 0.1, 0, 1)
    perturbation = adversarial - original

    # 測試單個攻擊視覺化
    print("\n測試單個攻擊視覺化...")
    visualize_attack(
        original, adversarial, perturbation,
        original_label=7, orig_pred=7, adv_pred=3,
        epsilon=0.3,
        save_path='test_single_attack.png'
    )

    # 測試準確率曲線
    print("\n測試準確率曲線...")
    epsilons = [0.0, 0.1, 0.2, 0.3]
    accuracies = [98.5, 85.2, 65.3, 45.8]
    plot_accuracy_vs_epsilon(epsilons, accuracies, 'test_accuracy_curve.png')

    # 測試攻擊成功率曲線
    print("\n測試攻擊成功率曲線...")
    success_rates = [0.0, 15.3, 35.2, 54.7]
    plot_attack_success_rate(epsilons, success_rates, 'test_success_rate.png')

    print("\n視覺化模組測試完成!")