Arko007 commited on
Commit
c47b969
·
verified ·
1 Parent(s): 5c957c9

Upload research_results.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. research_results.json +130 -0
research_results.json ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "methodology_note": "RQ1 (calibration) computed on the FULL held-out validation split (11666 frames / 1049 videos), 20-pass MC-Dropout (the plugin's designed default). RQ2/RQ3 (explanation-quality correlation, triage transferability) computed on a class-balanced stratified sample (up to 2000 real + 2000 fake frames per model, drawn from the same val split) because the perturbation-based explanation-stability metric requires one forward+backward pass per sample and is not batchable -- full-scale would be many hours of single-image GPU work per model. num_perturbations=3 matches backend/main.py's actual deployed setting exactly.",
3
+ "val_frame_count": 11666,
4
+ "val_video_count": 1049,
5
+ "models": {
6
+ "xception": {
7
+ "rq1_calibration": {
8
+ "frame_ece": 0.032874380892149876,
9
+ "video_ece": 0.037745453723143835,
10
+ "frame_auroc": 0.9976220768807692,
11
+ "video_auroc": 0.9995995550611791,
12
+ "frame_reliability": {
13
+ "bin_accuracies": [
14
+ 0.9478943611705924,
15
+ 0.15151515151515152,
16
+ 0.09090909090909091,
17
+ 0.044444444444444446,
18
+ 0.0625,
19
+ 0.9193548387096774,
20
+ 0.9393939393939394,
21
+ 1.0,
22
+ 1.0,
23
+ 0.9993681283963098
24
+ ],
25
+ "bin_confidences": [
26
+ 0.9961150191419919,
27
+ 0.8451707138482368,
28
+ 0.7526288335973567,
29
+ 0.6554717653327518,
30
+ 0.5481232423335314,
31
+ 0.552825871013826,
32
+ 0.6581012985923074,
33
+ 0.7533314636973447,
34
+ 0.8542808221161731,
35
+ 0.9956830007063067
36
+ ],
37
+ "bin_counts": [
38
+ 2802,
39
+ 66,
40
+ 55,
41
+ 45,
42
+ 48,
43
+ 62,
44
+ 66,
45
+ 86,
46
+ 163,
47
+ 7913
48
+ ],
49
+ "bin_boundaries": [
50
+ 0.0,
51
+ 0.1,
52
+ 0.2,
53
+ 0.30000000000000004,
54
+ 0.4,
55
+ 0.5,
56
+ 0.6000000000000001,
57
+ 0.7000000000000001,
58
+ 0.8,
59
+ 0.9,
60
+ 1.0
61
+ ]
62
+ },
63
+ "video_reliability": {
64
+ "bin_accuracies": [
65
+ 0.9733333333333334,
66
+ 0.4,
67
+ 1.0,
68
+ 0.3333333333333333,
69
+ 0.0,
70
+ 1.0,
71
+ 1.0,
72
+ 1.0,
73
+ 1.0,
74
+ 1.0
75
+ ],
76
+ "bin_confidences": [
77
+ 0.9954425858807838,
78
+ 0.8405209202261176,
79
+ 0.7561018553096801,
80
+ 0.6423113923558654,
81
+ 0.5399914891505614,
82
+ 0.5524718514084817,
83
+ 0.6568172973223576,
84
+ 0.7598564254777798,
85
+ 0.8647716873423498,
86
+ 0.9896538321624408
87
+ ],
88
+ "bin_counts": [
89
+ 150,
90
+ 5,
91
+ 1,
92
+ 3,
93
+ 5,
94
+ 5,
95
+ 13,
96
+ 34,
97
+ 54,
98
+ 777
99
+ ],
100
+ "bin_boundaries": [
101
+ 0.0,
102
+ 0.1,
103
+ 0.2,
104
+ 0.30000000000000004,
105
+ 0.4,
106
+ 0.5,
107
+ 0.6000000000000001,
108
+ 0.7000000000000001,
109
+ 0.8,
110
+ 0.9,
111
+ 1.0
112
+ ]
113
+ }
114
+ },
115
+ "rq2_uncertainty_explanation_correlation": {
116
+ "n_samples": 4000,
117
+ "n_skipped_timeout_or_error": 0,
118
+ "spearman_rho": -0.05301441006945229,
119
+ "p_value": 0.000795843020519685
120
+ },
121
+ "rq3_triage_transferability": {
122
+ "n_samples": 4000,
123
+ "n_false_negatives": 83,
124
+ "false_negative_escalation_rate": 0.9518072289156626,
125
+ "overall_escalation_rate": 0.65725,
126
+ "capture_ratio": 1.4481661908188095
127
+ }
128
+ }
129
+ }
130
+ }