Implementation:Norrrrrrr lyn WAInjectBench image ensemble detect

Knowledge Sources	WAInjectBench
Domains	Ensemble_Learning, Computer_Vision, Security
Last Updated	2026-02-14 16:00 GMT

Overview

Concrete tool for aggregating image detection results via set union across all detector outputs, provided by the WAInjectBench detector_image.ensemble module.

Description

The detect function in detector_image/ensemble.py is structurally identical to the text ensemble. It reads all *.jsonl files in the image result directory (excluding ensemble.jsonl), groups entries by data_name (folder name), unions detect_ids, and recomputes TPR/FPR.

Usage

Called when --detector ensemble is passed to main_image.py. Individual image detector results must already exist in the result directory.

Code Reference

Source Location

• Repository: WAInjectBench
• File: detector_image/ensemble.py (L7-60)

Signature

def detect(result_dir: str) -> List[Dict]:
    """
    Ensemble detector:
    - Reads all JSONL results from detectors in result_dir
    - Combines detect_ids by union
    - Recomputes TPR/FPR
    - Returns a list of results (same format as other detectors)
    """
    data_map = defaultdict(lambda: {"detect_ids": set(), "total_num": 0, "is_malicious": None})

    for file in Path(result_dir).glob("*.jsonl"):
        if file.name == "ensemble.jsonl":
            continue
        with open(file, "r", encoding="utf-8") as fin:
            for line in fin:
                entry = json.loads(line)
                data_name = entry["data_name"]
                data_map[data_name]["detect_ids"].update(entry.get("detect_ids", []))
                if data_map[data_name]["total_num"] == 0:
                    data_map[data_name]["total_num"] = entry.get("total_num", 0)
                if "tpr" in entry:
                    data_map[data_name]["is_malicious"] = True
                elif "fpr" in entry:
                    data_map[data_name]["is_malicious"] = False

    results = []
    for data_name, info in data_map.items():
        detect_ids = list(info["detect_ids"])
        total_num = info["total_num"]
        is_malicious = info["is_malicious"]
        if total_num > 0:
            if is_malicious:
                rate_key, rate_value = "tpr", round(len(detect_ids) / total_num, 4)
            else:
                rate_key, rate_value = "fpr", round(len(detect_ids) / total_num, 4)
        else:
            rate_key, rate_value = "tpr", 0.0 if is_malicious else "fpr", 0.0
        results.append({
            "data_name": data_name,
            rate_key: rate_value,
            "detect_ids": detect_ids,
            "total_num": total_num,
        })
    return results

Import

from detector_image import ensemble

I/O Contract

Inputs

Name	Type	Required	Description
result_dir	str	Yes	Directory containing per-detector .jsonl result files (default "result/image")

Outputs

Name	Type	Description
results	List[Dict]	Union-aggregated results with recomputed TPR/FPR

Usage Examples

Running Image Ensemble

from detector_image import ensemble

results = ensemble.detect("result/image")
for r in results:
    metric = "tpr" if "tpr" in r else "fpr"
    print(f"{r['data_name']}: {metric}={r[metric]}, detected={len(r['detect_ids'])}/{r['total_num']}")

Related Pages

Implements Principle

• Principle:Norrrrrrr_lyn_WAInjectBench_Ensemble_Aggregation_Image

Uses Heuristic

• Heuristic:Norrrrrrr_lyn_WAInjectBench_Union_Ensemble_Maximize_Recall