""" Phase 7 V2 Corrected External Generalization Evaluation Pipeline Runner Evaluates all three deepfake detectors with verified reference preprocessing: - ConvNeXt: DRCT native CenterCrop(224) without resizing (--is_crop configuration) - UnivFD: UniversalFakeDetect validate.py native CenterCrop(224) with CLIP normalization - ViT: HuggingFace ViTImageProcessor (verified bit-for-bit unchanged) Saves all outputs in benchmark/results/phase7_v2_reference_preprocessing/: - phase7_v2_vit_predictions.csv - phase7_v2_univfd_predictions.csv - phase7_v2_convnext_predictions.csv - phase7_v2_combined_predictions.csv - phase7_v2_combined_metrics.json - phase7_v2_model_comparison.csv Memory Safety: - Sequential child process isolation per detector guarantees 100% OS virtual memory reclamation. """ import os import sys import json import csv import time import math import argparse import subprocess import tempfile import gc from typing import Dict, Any, List, Tuple, Optional PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) if PROJECT_ROOT not in sys.path: sys.path.insert(0, PROJECT_ROOT) import torch import numpy as np from sklearn.metrics import roc_auc_score from src.detector_registry import ( get_detector, get_detector_info, reset_instances, ) from src.detector_schema import ( DetectorResult, PREDICTION_REAL, PREDICTION_FAKE, PREDICTION_ERROR, PREDICTION_UNCERTAIN, ) from src.decision_fusion import ( strategy_majority_vote, strategy_mean_fake_probability, strategy_vit_priority, ) MANIFEST_PATH = os.path.join(PROJECT_ROOT, "benchmark", "external_generalization_manifest.csv") OUT_DIR = os.path.join(PROJECT_ROOT, "benchmark", "results", "phase7_v2_reference_preprocessing") OUTPUT_VIT_CSV = os.path.join(OUT_DIR, "phase7_v2_vit_predictions.csv") OUTPUT_UNIVFD_CSV = os.path.join(OUT_DIR, "phase7_v2_univfd_predictions.csv") OUTPUT_CONVNEXT_CSV = os.path.join(OUT_DIR, "phase7_v2_convnext_predictions.csv") OUTPUT_COMBINED_CSV = os.path.join(OUT_DIR, "phase7_v2_combined_predictions.csv") OUTPUT_METRICS_JSON = os.path.join(OUT_DIR, "phase7_v2_combined_metrics.json") OUTPUT_COMPARISON_CSV = os.path.join(OUT_DIR, "phase7_v2_model_comparison.csv") V1_METRICS_JSON = os.path.join(PROJECT_ROOT, "benchmark", "results", "phase7_combined_metrics.json") V1_COMPARISON_CSV = os.path.join(PROJECT_ROOT, "benchmark", "results", "phase7_model_comparison.csv") DETECTOR_KEYS = ["vit", "univfd", "chen2024_convnext"] def wilson_score_interval(k: int, n: int, confidence: float = 0.95) -> List[float]: """Computes Wilson score confidence interval for binomial proportion k / n.""" if n <= 0: return [0.0, 0.0] p_hat = k / n z = 1.959963984540054 z2 = z * z denom = 1.0 + z2 / n center = (p_hat + z2 / (2.0 * n)) / denom margin = z * math.sqrt((p_hat * (1.0 - p_hat) / n) + (z2 / (4.0 * n * n))) / denom lower = max(0.0, center - margin) upper = min(1.0, center + margin) return [round(lower, 4), round(upper, 4)] def compute_cohen_kappa(both_fake: int, both_real: int, m1_f_m2_r: int, m1_r_m2_f: int) -> float: """Computes Cohen's Kappa for 2x2 agreement matrix.""" n = both_fake + both_real + m1_f_m2_r + m1_r_m2_f if n == 0: return 0.0 po = (both_fake + both_real) / n p_f1 = (both_fake + m1_f_m2_r) / n p_r1 = (m1_r_m2_f + both_real) / n p_f2 = (both_fake + m1_r_m2_f) / n p_r2 = (m1_f_m2_r + both_real) / n pe = (p_f1 * p_f2) + (p_r1 * p_r2) if abs(1.0 - pe) < 1e-9: return 1.0 if abs(po - 1.0) < 1e-9 else 0.0 kappa = (po - pe) / (1.0 - pe) return round(kappa, 4) def run_worker(det_key: str, input_json_path: str, output_json_path: str) -> None: """Worker function executed in isolated child process for a single detector.""" print(f" [Worker Process] Initializing detector '{det_key}' in clean subprocess...") with open(input_json_path, "r", encoding="utf-8") as f: samples = json.load(f) info = get_detector_info(det_key) meta = info.get("metadata", {}) fallback_det_name = meta.get("detector_name", det_key) fallback_model_name = meta.get("model_name", "Unknown") adapter = None load_err = None t0 = time.perf_counter() try: adapter = get_detector(det_key, lazy_load=False, singleton=False) load_time_sec = round(time.perf_counter() - t0, 3) print(f" [Worker Process] Detector '{det_key}' loaded successfully in {load_time_sec}s.") except Exception as e: load_err = str(e) load_time_sec = round(time.perf_counter() - t0, 3) print(f" [Worker Process] ERROR loading detector '{det_key}': {load_err}") results = [] latencies = [] for idx, sample in enumerate(samples, 1): file_path = sample["file_path"] sample_id = sample["sample_id"] sample_t0 = time.perf_counter() if load_err is not None or adapter is None: err_res = DetectorResult.create_error( detector_name=fallback_det_name, model_name=fallback_model_name, error_message=f"Model initialization failed: {load_err}", ) res_dict = err_res.to_dict() sample_latency_ms = round((time.perf_counter() - sample_t0) * 1000.0, 2) elif not os.path.exists(file_path): err_res = DetectorResult.create_error( detector_name=fallback_det_name, model_name=fallback_model_name, error_message=f"File not found on disk: {file_path}", ) res_dict = err_res.to_dict() sample_latency_ms = round((time.perf_counter() - sample_t0) * 1000.0, 2) else: try: res = adapter.predict(file_path) sample_latency_ms = round((time.perf_counter() - sample_t0) * 1000.0, 2) res_dict = res.to_dict() except Exception as inf_exc: sample_latency_ms = round((time.perf_counter() - sample_t0) * 1000.0, 2) err_res = DetectorResult.create_error( detector_name=fallback_det_name, model_name=fallback_model_name, error_message=f"Inference exception: {str(inf_exc)}", ) res_dict = err_res.to_dict() res_dict["latency_ms"] = sample_latency_ms latencies.append(sample_latency_ms) record = { "sample_id": sample_id, "detector_key": det_key, **res_dict } results.append(record) if idx % 50 == 0 or idx == len(samples): print(f" [Worker Process] Progress: {idx}/{len(samples)} samples processed.") # Explicit cleanup if adapter is not None and hasattr(adapter, "unload_model"): try: adapter.unload_model() except Exception: pass del adapter reset_instances() gc.collect() if torch.cuda.is_available(): torch.cuda.empty_cache() with open(output_json_path, "w", encoding="utf-8") as f: json.dump({ "results": results, "load_time_sec": load_time_sec, "mean_latency_ms": round(sum(latencies) / len(latencies), 2) if latencies else 0.0 }, f, indent=2) print(f" [Worker Process] Successfully completed all {len(results)} samples for '{det_key}'.") def run_detector_subprocess(det_key: str, samples: List[Dict[str, Any]]) -> Tuple[List[Dict[str, Any]], float, float]: """Runs a detector in an isolated child process to guarantee OS-level virtual memory cleanup.""" with tempfile.NamedTemporaryFile("w", encoding="utf-8", suffix=".json", delete=False) as f_in: json.dump(samples, f_in) in_path = f_in.name with tempfile.NamedTemporaryFile("w", encoding="utf-8", suffix=".json", delete=False) as f_out: out_path = f_out.name cmd = [ sys.executable, os.path.abspath(__file__), "--worker", "--worker-detector", det_key, "--worker-input", in_path, "--worker-output", out_path, ] try: proc = subprocess.run(cmd, capture_output=True, text=True, check=True) print(proc.stdout.strip()) with open(out_path, "r", encoding="utf-8") as f: data = json.load(f) return data["results"], data["load_time_sec"], data["mean_latency_ms"] except subprocess.CalledProcessError as exc: print(f" [ERROR] Subprocess for {det_key} failed with return code {exc.returncode}") print(" --- Subprocess STDOUT ---") print(exc.stdout) print(" --- Subprocess STDERR ---") print(exc.stderr) raise RuntimeError(f"Detector evaluation subprocess failed for {det_key}") from exc finally: for p in [in_path, out_path]: if os.path.exists(p): try: os.remove(p) except Exception: pass def compute_binary_metrics( predictions: List[str], ground_truths: List[str], probabilities: Optional[List[float]] = None ) -> Dict[str, Any]: """Computes binary classification metrics, Wilson CIs, and ROC-AUC.""" tp = sum(1 for p, g in zip(predictions, ground_truths) if p == PREDICTION_FAKE and g == PREDICTION_FAKE) fp = sum(1 for p, g in zip(predictions, ground_truths) if p == PREDICTION_FAKE and g == PREDICTION_REAL) tn = sum(1 for p, g in zip(predictions, ground_truths) if p == PREDICTION_REAL and g == PREDICTION_REAL) fn = sum(1 for p, g in zip(predictions, ground_truths) if p == PREDICTION_REAL and g == PREDICTION_FAKE) error_or_uncertain = sum(1 for p in predictions if p not in (PREDICTION_REAL, PREDICTION_FAKE)) total = len(predictions) valid_total = tp + fp + tn + fn actual_pos = tp + fn actual_neg = tn + fp accuracy = round((tp + tn) / valid_total, 4) if valid_total > 0 else 0.0 accuracy_ci = wilson_score_interval(tp + tn, valid_total) if valid_total > 0 else [0.0, 0.0] recall = round(tp / actual_pos, 4) if actual_pos > 0 else 0.0 recall_ci = wilson_score_interval(tp, actual_pos) if actual_pos > 0 else [0.0, 0.0] specificity = round(tn / actual_neg, 4) if actual_neg > 0 else 0.0 specificity_ci = wilson_score_interval(tn, actual_neg) if actual_neg > 0 else [0.0, 0.0] precision = round(tp / (tp + fp), 4) if (tp + fp) > 0 else 0.0 precision_ci = wilson_score_interval(tp, tp + fp) if (tp + fp) > 0 else [0.0, 0.0] balanced_acc = round(0.5 * (recall + specificity), 4) if (precision + recall) > 0: f1 = round(2.0 * precision * recall / (precision + recall), 4) else: f1 = 0.0 roc_auc = None if probabilities is not None and len(probabilities) == len(ground_truths) and actual_pos > 0 and actual_neg > 0: try: y_binary = [1 if g == PREDICTION_FAKE else 0 for g in ground_truths] valid_pairs = [(y, p) for y, p in zip(y_binary, probabilities) if p is not None] if len(valid_pairs) == len(ground_truths): roc_auc = round(float(roc_auc_score([y for y, _ in valid_pairs], [p for _, p in valid_pairs])), 4) except Exception: roc_auc = None return { "total_samples": total, "valid_evaluations": valid_total, "non_decisive_or_error": error_or_uncertain, "confusion_matrix": {"tp": tp, "fp": fp, "tn": tn, "fn": fn}, "accuracy": accuracy, "accuracy_ci_95": accuracy_ci, "balanced_accuracy": balanced_acc, "precision": precision, "precision_ci_95": precision_ci, "recall": recall, "recall_ci_95": recall_ci, "specificity": specificity, "specificity_ci_95": specificity_ci, "f1_score": f1, "roc_auc": roc_auc, } def main(): parser = argparse.ArgumentParser(description="Phase 7 V2 Corrected Evaluation Pipeline") parser.add_argument("--worker", action="store_true", help="Internal worker mode") parser.add_argument("--worker-detector", type=str, help="Detector key for worker") parser.add_argument("--worker-input", type=str, help="Input JSON for worker") parser.add_argument("--worker-output", type=str, help="Output JSON for worker") parser.add_argument("--skip-smoke", action="store_true", help="Skip preliminary smoke tests") parser.add_argument("--force-rerun", action="store_true", help="Force re-running inference even if CSVs exist") args = parser.parse_args() if args.worker: run_worker(args.worker_detector, args.worker_input, args.worker_output) return print("=" * 80) print("PHASE 7 V2: CORRECTED EXTERNAL GENERALIZATION EVALUATION") print("Official Reference Preprocessing Alignment Verification") print("=" * 80) os.makedirs(OUT_DIR, exist_ok=True) # 1. Load manifest print(f"Loading external generalization manifest from: {MANIFEST_PATH}") if not os.path.exists(MANIFEST_PATH): raise FileNotFoundError(f"Manifest not found: {MANIFEST_PATH}") samples = [] with open(MANIFEST_PATH, "r", encoding="utf-8") as f: reader = csv.DictReader(f) for row in reader: samples.append({ "sample_id": row["sample_id"], "original_filename": row["original_filename"], "file_path": row["absolute_path"], "ground_truth": row["label"], "generator_family": row["generator_family"], "category": row.get("category", ""), }) total_samples = len(samples) real_count = sum(1 for s in samples if s["ground_truth"] == PREDICTION_REAL) fake_count = sum(1 for s in samples if s["ground_truth"] == PREDICTION_FAKE) midjourney_count = sum(1 for s in samples if s["generator_family"] == "midjourney") adm_count = sum(1 for s in samples if s["generator_family"] == "adm") print(f"Manifest verified: {total_samples} samples total (REAL={real_count}, FAKE={fake_count} [Midjourney={midjourney_count}, ADM={adm_count}])") assert total_samples == 200, f"Expected 200 samples, got {total_samples}" assert real_count == 100, f"Expected 100 REAL samples, got {real_count}" assert fake_count == 100, f"Expected 100 FAKE samples, got {fake_count}" # 2. Preliminary Smoke Tests if not args.skip_smoke and not (os.path.exists(OUTPUT_VIT_CSV) and not args.force_rerun): print("\n--- Step 1: Preliminary Single-Sample Smoke Tests ---") smoke_sample = [samples[0]] for det_key in DETECTOR_KEYS: print(f"Testing detector '{det_key}' smoke execution...") smoke_res, load_t, lat_t = run_detector_subprocess(det_key, smoke_sample) print(f" -> '{det_key}' smoke passed! Pred: {smoke_res[0].get('prediction')}, Latency: {lat_t}ms") # 3. Sequential Subprocess Inference across 200 samples print("\n--- Step 2: Sequential Detector Inference on 200 Samples ---") detector_outputs: Dict[str, List[Dict[str, Any]]] = {} detector_perf: Dict[str, Dict[str, float]] = {} all_csvs_exist = all(os.path.exists(p) for p in [OUTPUT_VIT_CSV, OUTPUT_UNIVFD_CSV, OUTPUT_CONVNEXT_CSV]) if all_csvs_exist and not args.force_rerun: print("Reusing existing completed individual predictions from Phase 7 V2 CSVs...") for det_key, out_path in [ ("vit", OUTPUT_VIT_CSV), ("univfd", OUTPUT_UNIVFD_CSV), ("chen2024_convnext", OUTPUT_CONVNEXT_CSV), ]: rows = [] with open(out_path, "r", encoding="utf-8") as f: for r in csv.DictReader(f): rows.append({ "sample_id": r["sample_id"], "detector_key": det_key, "detector_name": r.get("detector_name", ""), "model_name": r.get("model_name", ""), "prediction": r.get("prediction", ""), "predicted_class": r.get("predicted_class", r.get("prediction", "")), "fake_probability": float(r["fake_probability"]) if r.get("fake_probability") else None, "real_probability": float(r["real_probability"]) if r.get("real_probability") else None, "confidence": float(r["confidence"]) if r.get("confidence") else None, "raw_logit": float(r["raw_logit"]) if r.get("raw_logit") else None, "threshold": float(r.get("threshold", 0.5)), "latency_ms": float(r.get("latency_ms", 0.0)) if r.get("latency_ms") else 0.0, "error": r.get("error_message"), "status": r.get("status"), "preprocessing": r.get("preprocessing"), }) detector_outputs[det_key] = rows latencies = [r["latency_ms"] for r in rows if r["latency_ms"] > 0] detector_perf[det_key] = { "load_time_sec": 0.0, "mean_latency_ms": round(sum(latencies) / len(latencies), 2) if latencies else 0.0, "total_time_sec": 0.0, } print(f" Loaded {len(rows)} samples for '{det_key}' from {out_path} (Mean Latency: {detector_perf[det_key]['mean_latency_ms']}ms)") else: for det_key in DETECTOR_KEYS: print(f"\nEvaluating detector '{det_key}' on 200 samples...") t_start = time.perf_counter() results, load_time, mean_lat = run_detector_subprocess(det_key, samples) t_total = time.perf_counter() - t_start print(f"Completed '{det_key}' in {t_total:.2f}s (Load: {load_time}s, Mean Latency: {mean_lat}ms)") detector_outputs[det_key] = results detector_perf[det_key] = { "load_time_sec": load_time, "mean_latency_ms": mean_lat, "total_time_sec": round(t_total, 2), } # 4. Save Individual Prediction CSVs print("\n--- Step 3: Exporting Individual Model Predictions (Phase 7 V2) ---") pred_fieldnames = [ "sample_id", "original_filename", "ground_truth", "generator_family", "detector_key", "detector_name", "model_name", "preprocessing_version", "preprocessing", "prediction", "predicted_class", "fake_probability", "real_probability", "confidence", "raw_logit", "threshold", "status", "error_message", "latency_ms", "file_path" ] preprocessing_versions = { "vit": "Phase 7 v2 (HuggingFace AutoImageProcessor, unchanged)", "univfd": "Phase 7 v2 (Official validate.py native CenterCrop 224, no resize)", "chen2024_convnext": "Phase 7 v2 (Official DRCT --is_crop native CenterCrop 224, no resize)", } for det_key, out_path in [ ("vit", OUTPUT_VIT_CSV), ("univfd", OUTPUT_UNIVFD_CSV), ("chen2024_convnext", OUTPUT_CONVNEXT_CSV), ]: res_list = detector_outputs[det_key] with open(out_path, "w", encoding="utf-8", newline="") as f: writer = csv.DictWriter(f, fieldnames=pred_fieldnames) writer.writeheader() for r, s in zip(res_list, samples): row_dict = { "sample_id": s["sample_id"], "original_filename": s["original_filename"], "ground_truth": s["ground_truth"], "generator_family": s["generator_family"], "detector_key": det_key, "detector_name": r.get("detector_name", ""), "model_name": r.get("model_name", ""), "preprocessing_version": preprocessing_versions[det_key], "preprocessing": r.get("preprocessing", ""), "prediction": r.get("prediction", ""), "predicted_class": r.get("predicted_class", r.get("prediction", "")), "fake_probability": r.get("fake_probability", ""), "real_probability": r.get("real_probability", ""), "confidence": r.get("confidence", ""), "raw_logit": r.get("raw_logit", ""), "threshold": r.get("threshold", 0.5), "status": r.get("status", ""), "error_message": r.get("error", "") or r.get("error_message", ""), "latency_ms": r.get("latency_ms", ""), "file_path": s["file_path"], } writer.writerow(row_dict) print(f" Written {len(res_list)} rows to: {out_path}") # 5. Combined Classifier & Ensemble Predictions print("\n--- Step 4: Running Decision Fusion Across All Samples ---") vit_preds = [r.get("prediction") for r in detector_outputs["vit"]] univfd_preds = [r.get("prediction") for r in detector_outputs["univfd"]] convnext_preds = [r.get("prediction") for r in detector_outputs["chen2024_convnext"]] vit_probs = [r.get("fake_probability") for r in detector_outputs["vit"]] univfd_probs = [r.get("fake_probability") for r in detector_outputs["univfd"]] convnext_probs = [r.get("fake_probability") for r in detector_outputs["chen2024_convnext"]] combined_rows = [] maj_preds = [] maj_scores = [] # continuous score: fake vote fraction (0.0, 1/3, 2/3, 1.0) mean_preds = [] mean_probs = [] # continuous score: mean fake probability vit_priority_preds = [] vit_priority_probs = [] # continuous score: vit fake probability agree_3_count = 0 for i, s in enumerate(samples): record_map = { "vit": detector_outputs["vit"][i], "univfd": detector_outputs["univfd"][i], "chen2024_convnext": detector_outputs["chen2024_convnext"][i], } # 1. Majority Voting # Calculate vote fraction as true continuous score valid_votes = [vit_preds[i], univfd_preds[i], convnext_preds[i]] fake_votes = sum(1 for v in valid_votes if v == PREDICTION_FAKE) maj_score = round(fake_votes / len(valid_votes), 6) res_maj = strategy_majority_vote(record_map, tie_breaker="vit_or_real") maj_preds.append(res_maj.prediction) maj_scores.append(maj_score) # 2. Mean Fake Probability res_mean = strategy_mean_fake_probability(record_map, threshold=0.50) mean_preds.append(res_mean.prediction) mean_probs.append(res_mean.fake_probability) # 3. ViT-Priority Fusion res_prio = strategy_vit_priority(record_map, vit_key="vit") vit_priority_preds.append(res_prio.prediction) vit_priority_probs.append(res_prio.fake_probability) all_three_agree = (vit_preds[i] == univfd_preds[i] == convnext_preds[i]) if all_three_agree: agree_3_count += 1 combined_rows.append({ "sample_id": s["sample_id"], "original_filename": s["original_filename"], "ground_truth": s["ground_truth"], "generator_family": s["generator_family"], "vit_pred": vit_preds[i], "vit_fake_prob": vit_probs[i], "univfd_pred": univfd_preds[i], "univfd_fake_prob": univfd_probs[i], "convnext_pred": convnext_preds[i], "convnext_fake_prob": convnext_probs[i], "majority_vote_pred": res_maj.prediction, "majority_vote_score": maj_score, "majority_vote_fake_prob": res_maj.fake_probability, "mean_prob_pred": res_mean.prediction, "mean_prob_fake_prob": res_mean.fake_probability, "vit_priority_pred": res_prio.prediction, "vit_priority_fake_prob": res_prio.fake_probability, "all_three_agree": all_three_agree, "majority_explanation": res_maj.explanation, "mean_explanation": res_mean.explanation, "priority_explanation": res_prio.explanation, }) with open(OUTPUT_COMBINED_CSV, "w", encoding="utf-8", newline="") as f: writer = csv.DictWriter(f, fieldnames=list(combined_rows[0].keys())) writer.writeheader() writer.writerows(combined_rows) print(f" Written combined predictions ({len(combined_rows)} rows) to: {OUTPUT_COMBINED_CSV}") # 6. Compute Comprehensive Metrics print("\n--- Step 5: Computing Statistical Metrics (Aggregate & Per-Generator) ---") ground_truths = [s["ground_truth"] for s in samples] targets = { "vit": { "name": "ViT (dima806)", "preds": vit_preds, "scores": vit_probs, "lat": detector_perf["vit"]["mean_latency_ms"] }, "univfd": { "name": "UnivFD (CLIP ViT-L/14)", "preds": univfd_preds, "scores": univfd_probs, "lat": detector_perf["univfd"]["mean_latency_ms"] }, "chen2024_convnext": { "name": "ConvNeXt (DRCT / RAID)", "preds": convnext_preds, "scores": convnext_probs, "lat": detector_perf["chen2024_convnext"]["mean_latency_ms"] }, "ensemble_majority_vote": { "name": "Ensemble (Majority Vote)", "preds": maj_preds, "scores": maj_scores, # evaluated using vote fraction continuous score! "lat": sum(detector_perf[k]["mean_latency_ms"] for k in DETECTOR_KEYS) }, "ensemble_mean_prob": { "name": "Ensemble (Mean Fake Probability)", "preds": mean_preds, "scores": mean_probs, "lat": sum(detector_perf[k]["mean_latency_ms"] for k in DETECTOR_KEYS) }, "ensemble_vit_priority": { "name": "Ensemble (ViT Priority Fallback)", "preds": vit_priority_preds, "scores": vit_priority_probs, "lat": detector_perf["vit"]["mean_latency_ms"] }, } # Slice indices idx_real = [i for i, s in enumerate(samples) if s["ground_truth"] == PREDICTION_REAL] idx_mj = [i for i, s in enumerate(samples) if s["generator_family"] == "midjourney"] idx_adm = [i for i, s in enumerate(samples) if s["generator_family"] == "adm"] # Slice cohorts (authentic controls + generator synthetic) idx_mj_cohort = idx_real + idx_mj idx_adm_cohort = idx_real + idx_adm aggregate_results: Dict[str, Any] = {} generator_slice_results: Dict[str, Any] = {} for k, t in targets.items(): # Aggregate metrics agg_m = compute_binary_metrics(t["preds"], ground_truths, t["scores"]) agg_m["mean_latency_ms"] = t["lat"] agg_m["display_name"] = t["name"] aggregate_results[k] = agg_m # Midjourney cohort (100 Real + 50 Midjourney = 150 samples) mj_preds = [t["preds"][i] for i in idx_mj_cohort] mj_gts = [ground_truths[i] for i in idx_mj_cohort] mj_scores = [t["scores"][i] for i in idx_mj_cohort] mj_m = compute_binary_metrics(mj_preds, mj_gts, mj_scores) # ADM cohort (100 Real + 50 ADM = 150 samples) adm_preds = [t["preds"][i] for i in idx_adm_cohort] adm_gts = [ground_truths[i] for i in idx_adm_cohort] adm_scores = [t["scores"][i] for i in idx_adm_cohort] adm_m = compute_binary_metrics(adm_preds, adm_gts, adm_scores) # Standalone fake detections (pure recall out of 50) mj_raw_correct = sum(1 for i in idx_mj if t["preds"][i] == PREDICTION_FAKE) adm_raw_correct = sum(1 for i in idx_adm if t["preds"][i] == PREDICTION_FAKE) real_raw_correct = sum(1 for i in idx_real if t["preds"][i] == PREDICTION_REAL) # Score stats on slices raw_scores = t["scores"] real_scores = [raw_scores[i] for i in idx_real if raw_scores[i] is not None] mj_scores_only = [raw_scores[i] for i in idx_mj if raw_scores[i] is not None] adm_scores_only = [raw_scores[i] for i in idx_adm if raw_scores[i] is not None] generator_slice_results[k] = { "display_name": t["name"], "authentic_control": { "total": len(idx_real), "correct_real": real_raw_correct, "specificity": round(real_raw_correct / len(idx_real), 4), "score_mean": round(float(np.mean(real_scores)), 4) if real_scores else None, "score_std": round(float(np.std(real_scores)), 4) if real_scores else None, }, "midjourney_slice": { "total_cohort_samples": len(idx_mj_cohort), "accuracy": mj_m["accuracy"], "recall": mj_m["recall"], "specificity": mj_m["specificity"], "precision": mj_m["precision"], "f1_score": mj_m["f1_score"], "roc_auc": mj_m["roc_auc"], "raw_detections": f"{mj_raw_correct} / {len(idx_mj)} ({mj_raw_correct / len(idx_mj) * 100:.1f}%)", "score_mean": round(float(np.mean(mj_scores_only)), 4) if mj_scores_only else None, "score_std": round(float(np.std(mj_scores_only)), 4) if mj_scores_only else None, }, "adm_slice": { "total_cohort_samples": len(idx_adm_cohort), "accuracy": adm_m["accuracy"], "recall": adm_m["recall"], "specificity": adm_m["specificity"], "precision": adm_m["precision"], "f1_score": adm_m["f1_score"], "roc_auc": adm_m["roc_auc"], "raw_detections": f"{adm_raw_correct} / {len(idx_adm)} ({adm_raw_correct / len(idx_adm) * 100:.1f}%)", "score_mean": round(float(np.mean(adm_scores_only)), 4) if adm_scores_only else None, "score_std": round(float(np.std(adm_scores_only)), 4) if adm_scores_only else None, }, } # 7. Pairwise Agreement Matrix & Cohen's Kappa print("\n--- Step 6: Computing Pairwise Detector Agreement ---") pairwise_pairs = [ ("vit", "univfd"), ("vit", "chen2024_convnext"), ("univfd", "chen2024_convnext"), ] pairwise_analysis = {} for d1, d2 in pairwise_pairs: p1 = targets[d1]["preds"] p2 = targets[d2]["preds"] both_fake = sum(1 for a, b in zip(p1, p2) if a == PREDICTION_FAKE and b == PREDICTION_FAKE) both_real = sum(1 for a, b in zip(p1, p2) if a == PREDICTION_REAL and b == PREDICTION_REAL) d1_f_d2_r = sum(1 for a, b in zip(p1, p2) if a == PREDICTION_FAKE and b == PREDICTION_REAL) d1_r_d2_f = sum(1 for a, b in zip(p1, p2) if a == PREDICTION_REAL and b == PREDICTION_FAKE) agree_count = both_fake + both_real agree_pct = round(agree_count / total_samples * 100.0, 2) kappa = compute_cohen_kappa(both_fake, both_real, d1_f_d2_r, d1_r_d2_f) pairwise_analysis[f"{d1}_vs_{d2}"] = { "detector_1": targets[d1]["name"], "detector_2": targets[d2]["name"], "agreed_samples": agree_count, "total_samples": total_samples, "agreement_percentage": agree_pct, "both_fake": both_fake, "both_real": both_real, "d1_fake_d2_real": d1_f_d2_r, "d1_real_d2_fake": d1_r_d2_f, "cohen_kappa": kappa, } print(f" {targets[d1]['name']} vs {targets[d2]['name']}: Agreement={agree_pct}% (Kappa={kappa})") # 8. Comparison Table: Phase 7 V1 vs Phase 7 V2 print("\n--- Step 7: Exporting Model Comparison CSV ---") comparison_rows = [] comp_fieldnames = [ "Model / Strategy", "Aggregate Acc (%)", "Aggregate 95% CI", "Aggregate BalAcc (%)", "Aggregate Precision", "Aggregate Recall", "Aggregate Specificity", "Aggregate F1", "Aggregate ROC-AUC", "Midjourney Acc (%)", "Midjourney Recall (%)", "ADM Acc (%)", "ADM Recall (%)", "TP", "FP", "TN", "FN", "Mean Latency (ms)" ] for k, t in targets.items(): agg = aggregate_results[k] slices = generator_slice_results[k] cm = agg["confusion_matrix"] row = { "Model / Strategy": t["name"], "Aggregate Acc (%)": f"{agg['accuracy'] * 100:.2f}", "Aggregate 95% CI": f"[{agg['accuracy_ci_95'][0] * 100:.2f}%, {agg['accuracy_ci_95'][1] * 100:.2f}%]", "Aggregate BalAcc (%)": f"{agg['balanced_accuracy'] * 100:.2f}", "Aggregate Precision": f"{agg['precision']:.4f}", "Aggregate Recall": f"{agg['recall']:.4f}", "Aggregate Specificity": f"{agg['specificity']:.4f}", "Aggregate F1": f"{agg['f1_score']:.4f}", "Aggregate ROC-AUC": f"{agg['roc_auc']:.4f}" if agg['roc_auc'] is not None else "N/A", "Midjourney Acc (%)": f"{slices['midjourney_slice']['accuracy'] * 100:.2f}", "Midjourney Recall (%)": f"{slices['midjourney_slice']['recall'] * 100:.2f}", "ADM Acc (%)": f"{slices['adm_slice']['accuracy'] * 100:.2f}", "ADM Recall (%)": f"{slices['adm_slice']['recall'] * 100:.2f}", "TP": cm["tp"], "FP": cm["fp"], "TN": cm["tn"], "FN": cm["fn"], "Mean Latency (ms)": f"{agg['mean_latency_ms']:.1f}", } comparison_rows.append(row) with open(OUTPUT_COMPARISON_CSV, "w", encoding="utf-8", newline="") as f: writer = csv.DictWriter(f, fieldnames=comp_fieldnames) writer.writeheader() writer.writerows(comparison_rows) print(f" Written comparison table ({len(comparison_rows)} rows) to: {OUTPUT_COMPARISON_CSV}") # 9. Load Version 1 data for direct delta calculation v1_metrics = {} if os.path.exists(V1_METRICS_JSON): try: with open(V1_METRICS_JSON, "r", encoding="utf-8") as f: v1_metrics = json.load(f) except Exception: pass version_deltas = {} for k in targets.keys(): v2_acc = aggregate_results[k]["accuracy"] v2_rec = aggregate_results[k]["recall"] v2_spec = aggregate_results[k]["specificity"] v2_auc = aggregate_results[k]["roc_auc"] v1_acc = v1_metrics.get("aggregate_performance", {}).get(k, {}).get("accuracy") v1_rec = v1_metrics.get("aggregate_performance", {}).get(k, {}).get("recall") v1_spec = v1_metrics.get("aggregate_performance", {}).get(k, {}).get("specificity") v1_auc = v1_metrics.get("aggregate_performance", {}).get(k, {}).get("roc_auc") version_deltas[k] = { "v1_accuracy": v1_acc, "v2_accuracy": v2_acc, "delta_accuracy": round(v2_acc - v1_acc, 4) if v1_acc is not None else None, "v1_recall": v1_rec, "v2_recall": v2_rec, "delta_recall": round(v2_rec - v1_rec, 4) if v1_rec is not None else None, "v1_specificity": v1_spec, "v2_specificity": v2_spec, "delta_specificity": round(v2_spec - v1_spec, 4) if v1_spec is not None else None, "v1_roc_auc": v1_auc, "v2_roc_auc": v2_auc, "delta_roc_auc": round(v2_auc - v1_auc, 4) if (v1_auc is not None and v2_auc is not None) else None, } # 10. Save Complete Combined Metrics JSON metrics_payload = { "metadata": { "benchmark_phase": "Phase 7 Version 2 (Official Reference Preprocessing)", "timestamp": time.strftime("%Y-%m-%dT%H:%M:%SZ"), "manifest_path": "benchmark/external_generalization_manifest.csv", "dataset_source": "GenImage (TheKernel01/Tiny-GenImage validation split)", "decision_threshold": 0.50, "total_samples": 200, "real_samples": 100, "fake_samples": 100, "generator_counts": { "midjourney": 50, "adm": 50 }, "environment": { "python_version": sys.version, "torch_version": torch.__version__, "platform": sys.platform, }, }, "aggregate_performance": aggregate_results, "generator_slices": generator_slice_results, "pairwise_agreement": pairwise_analysis, "all_three_agree_count": agree_3_count, "all_three_agree_pct": round(agree_3_count / total_samples * 100.0, 2), "version_comparison_v1_vs_v2": version_deltas, } with open(OUTPUT_METRICS_JSON, "w", encoding="utf-8") as f: json.dump(metrics_payload, f, indent=2) print(f" Written complete metrics JSON to: {OUTPUT_METRICS_JSON}") print("\n" + "=" * 80) print("PHASE 7 V2 EVALUATION COMPLETE") print("=" * 80) for row in comparison_rows: print(f" {row['Model / Strategy']:35s} | Acc: {row['Aggregate Acc (%)']:>6s}% | Recall: {row['Aggregate Recall']:>6s} | Spec: {row['Aggregate Specificity']:>6s} | AUC: {row['Aggregate ROC-AUC']:>6s}") if __name__ == "__main__": main()