SISS Defense Protocol - Attack Success: leaderboard

SISS measures changes in hidden-text recovery between native and aggressively downsampled images. That cross-scale shift is distinct from a web agent taking an attacker-directed action. Main, web-agent and defense experiments use separate protocols; conditional remaining attack success uses the publisher’s paired counts. This board reports defended attack success rate (%). Lower is better.

Metric: Defended attack success rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1Qwen2.5-VL-7B-Instruct (Harness: SISS defense ScaleGuard (adaptive))15
2Qwen2.5-VL-7B-Instruct (Harness: SISS defense ScaleGuard (non-adaptive))24.3
3Qwen 2.5 VL 7B Instruct (Harness: SISS defense Area resampling)67
4Qwen 2.5 VL 7B Instruct (Harness: SISS defense Median reconstruction)67
5Qwen 2.5 VL 7B Instruct (Harness: SISS defense Gaussian low-pass)67
6Qwen2.5-VL-7B-Instruct (Harness: SISS defense Randomized masking (5 reps.))68
7Qwen2.5-VL-7B-Instruct (Harness: SISS defense JPEG(75) + + area)69

Interactive version: theaggregate.ai/benchmark?slug=siss-defense-protocol-attack-success · How It Works · Data refreshed daily, snapshot 2026-11-16.