Update index.html

index.html

@@ -806,6 +806,22 @@
                     </tbody>
                 </table>
                 
+                <h3>Screenshot Survival Benchmark</h3>
+                <p>To simulate a common method of bypassing protection, we ran a rigorous benchmark on a screenshot of a protected image. The test automatically aligns the original and screenshot images, measures the surviving poison, and re-evaluates its impact on ML models, including a full fine-tuning test.</p>
+                
+                <img src="assets/screenshot_visual_report_enhanced.png" alt="Screenshot Survival Benchmark Dashboard" style="width: 100%; border-radius: 12px; margin: 20px 0; border: 2px solid rgba(163, 120, 72, 0.3);">
+                
+                <div style="background: rgba(163, 120, 72, 0.15); padding: 20px; border-radius: 8px; margin: 20px 0; border-left: 4px solid var(--premium-gold);">
+                    <h4 style="margin-bottom: 15px; color: var(--premium-gold);">Key Results (Post-Screenshot):</h4>
+                    <ul style="line-height: 1.8;">
+                        <li><strong>✅ ML Training Disruption:</strong> <strong style="color: var(--success-color);">31.6% training degradation.</strong> The model trained on screenshot data had a 31.6% higher final loss, proving the surviving poison significantly hinders the learning process.</li>
+                        <li><strong>✅ Frequency Survival:</strong> 61.9% of the surviving armor's energy remains in the critical mid-band, demonstrating exceptional resilience.</li>
+                        <li><strong>✅ Perceptual Quality:</strong> SSIM of 0.822, indicating the image is still visually coherent after screenshotting.</li>
+                        <li><strong>ML Feature Degradation:</strong> While direct feature degradation was low (~1.2%), the far more critical fine-tuning test confirmed the armor's powerful real-world impact on model training.</li>
+                    </ul>
+                    <p style="margin-top: 15px; font-style: italic;"><strong>Conclusion:</strong> Poisonous Shield for Images survives the screenshot process and remains highly effective at poisoning the ML training pipeline—a critical feature for real-world creative protection.</p>
+                </div>
+
                 <h3>Visual Examples: Natural Photography</h3>
                 <p style="margin-bottom: 15px;">Park scene tested at different protection strengths:</p>