Datasets:

datamatters24
/

cano-adversarial-privacy

@@ -1,478 +0,0 @@
-================================================================================
-  CANO: Context-Aware Noise Optimization for Adversarial Privacy Protection
-================================================================================
-AUTHORS: Ted Rubin
-AFFILIATION: Independent Researcher
-EMAIL: ted@theorubin.com
-================================================================================
-ABSTRACT
-================================================================================
-We present CANO (Context-Aware Noise Optimization), an adaptive noise injection
-system that optimizes the privacy-utility tradeoff in adversarial privacy
-protection. Unlike uniform noise strategies, CANO allocates noise proportionally
-to each feature's contribution to re-identification, concentrating protection
-where it matters most while preserving utility on low-impact features.
-We evaluate CANO against five baseline strategies (Gaussian, FGSM, PGD,
-Carlini-Wagner, and Laplace) across 68,885 experimental configurations
-spanning 12 datasets (11 after excluding the 2-user cybersec_intrusion dataset
-from aggregate statistics), 3 attack models, and 6 noise budgets. Aggregate
-statistics are computed over 54,281 in-scope configurations, including a
-complete 540-row block on the real FP-Stalker browser-fingerprint corpus
-(Vastel et al. [10]; 776 users, 13,674 fingerprints, 34 attributes).
-Against a known adaptive attacker, CANO achieves a mean accuracy reduction of
-0.112 +/- 0.178 (see Fig. 1, Fig. 3) -- below
-Gaussian noise (0.395) but above C&W (0.001).
-However, two findings reframe this aggregate result. First, on the FP-Stalker
-real browser-fingerprint corpus the CANO/Gaussian gap collapses from the
-typical small-synthetic ratio of 100:1 (e.g., synth_50u_20s: CANO 0.003 vs
-Gaussian 0.514) to roughly 4:5 (CANO 0.276 vs Gaussian 0.340), suggesting
-that importance-weighted allocation is much closer to competitive when
-feature importance reflects realistic attribute redundancy rather than
-hand-crafted synthetic noise. Second, CANO achieves a strong transfer-attack
-profile: a transfer-to-adaptive ratio of 2.35x (transfer reduction 0.263 vs
-adaptive 0.112), compared to 1.04x for Gaussian. CANO's feature-importance
-allocation produces perturbations that generalize better across unseen
-attack models (see Fig. 2 Pareto front) -- the realistic deployment setting.
-CANO also produces lower noise magnitude than Gaussian/FGSM while retaining
-higher SNR.
-In adversarial co-evolutionary training, the DQN policy reduces attacker
-re-identification accuracy from 74.8% to 20.8% within 30 rounds (Fig. 6),
-converging to near-uniform allocation (Gini coefficient: 0.009) -- empirically
-demonstrating that uniform noise is the game-theoretic equilibrium against
-adaptive adversaries.
-Keywords: privacy protection, adversarial noise, reinforcement learning,
-          browser fingerprinting, context-aware optimization, transfer attacks
-================================================================================
-1. INTRODUCTION
-================================================================================
-Browser fingerprinting poses a significant threat to user privacy. Attackers
-construct unique device fingerprints from browser attributes -- canvas
-rendering, WebGL, screen resolution, installed fonts -- to track users across
-sessions without cookies [1].
-Privacy-preserving systems combat fingerprinting by injecting noise. A
-fundamental unresolved tension: whether uniform noise injection or
-feature-weighted injection provides superior protection. A further practical
-challenge: privacy systems are typically deployed without knowledge of the
-adversary's exact attack model.
-CANO addresses this through:
-  (1) Feature importance analysis.
-  (2) Proportional noise allocation with a minimum weight floor preventing
-      exploitable zero-noise features.
-  (3) RL that adapts allocation through adversarial co-evolution.
-  (4) Empirical analysis of the adaptive-vs-transfer tradeoff.
-Our central finding is counterintuitive: while CANO does not maximize accuracy
-reduction against a known adaptive attacker (Gaussian dominates), CANO achieves
-a 2.35x transfer-to-adaptive ratio -- notably higher than Gaussian's
-1.04x. In real deployments, defenders cannot tailor their noise to
-the adversary's model.
-1.1 Related Work
-Browser fingerprinting was popularized by Eckersley's Panopticlick study [7],
-which showed that the combination of routine browser attributes -- user
-agent, screen resolution, time zone, plugin list -- is enough to uniquely
-identify the vast majority of users. Subsequent work expanded the attack
-surface to canvas rendering, WebGL hashes, font metrics, and JavaScript-
-exposed APIs; Laperdrix et al.'s 2020 survey [1] catalogues 17 distinct
-categories of fingerprinting signals and remains the standard reference.
-FP-Stalker (Vastel et al. [10]) introduced longitudinal evaluation by
-tracking how individual fingerprints evolve over weeks, which is why we
-adopt its 776-user corpus as our reference real-data benchmark in
-Section 3.5. On the defensive side, BrFAST (Andriamilanto & Allard [8])
-and FPSelect [9] focus on attribute *selection* -- which fingerprint
-attributes a privacy-conscious browser should expose at all -- whereas
-CANO operates on a complementary axis: given that an attribute is exposed,
-how much per-attribute noise to inject and how to allocate that noise
-across the attribute set.
-The privacy-utility tradeoff via random perturbation has a long lineage in
-differential privacy (Dwork et al. [5]), where uniform additive Gaussian or
-Laplace noise is calibrated to a per-query sensitivity bound. The
-adversarial-examples literature complements this by showing that targeted
-perturbations can dramatically reduce classifier accuracy under L2 or
-L-infinity budgets: FGSM [2] is a single-step gradient attack, PGD [3]
-iterates it under projection, and Carlini-Wagner [4] casts the same
-problem as constrained optimization. CANO sits between these two
-traditions. It borrows the budget-controlled framing from adversarial
-examples but allocates the budget by a *static* feature-importance prior,
-rather than by per-input gradient (FGSM/PGD) or by uniform calibration
-(DP). The minimum-weight floor (min_weight = 0.1, Section 2.2) is a
-defensive concession to the same observation that motivates DP's worst-
-case framing: any feature that receives systematically less noise becomes
-the attacker's preferred discriminator.
-A defense's robustness to *adaptive* adversaries -- those that retrain on
-protected outputs rather than the clean distribution -- is typically much
-weaker than its initial efficacy [3]. We confirm this in our adversarial
-co-evolutionary training (Section 3.7): the DQN policy converges to a
-near-uniform allocation (Gini = 0.009 after 30 rounds), consistent with the
-intuition that under retraining pressure the optimal defense distributes
-noise broadly. The orthogonal axis is *transferability*: a perturbation
-crafted against one attacker model may or may not generalize to a different
-one (cross-model transfer is a long-standing topic in adversarial ML).
-Our central empirical contribution measures the transfer-vs-adaptive gap
-explicitly across all six strategies, and finds that CANO's importance-
-weighted allocation produces perturbations with a 2.35x transfer-to-
-adaptive ratio versus 1.04x for Gaussian -- the realistic deployment
-regime, since defenders typically do not know the deployed attack model
-and must rely on transfer.
-================================================================================
-2. METHODOLOGY
-================================================================================
-2.1 Feature Importance Analysis
-Random Forest (100 trees) on simulated fingerprint samples; permutation
-importance (10 repeats). 9-dimensional feature space of normalized browser
-fingerprint attributes. Baseline re-identification accuracy: 100% on synthetic
-corpus.
-Feature importance is highly concentrated: feature_0 (0.303) and feature_1
-(0.302) account for ~99% of total importance; features 2-5 have exactly 0.000
-permutation importance (Fig. 5). This motivates the minimum weight floor
-(min_weight = 0.1) in Section 2.2.
-[NOTE: Synthetic proxies, not real browser API measurements. Real-world
-validation is in progress against the FP-Stalker corpus (Vastel et al.,
-IEEE S&P 2018 [10]; ~21,809 real browser fingerprints across 40
-attributes) with per-attribute instability and memory costs from BrFAST
-(Andriamilanto & Allard, WWW '21 Companion [8]) supplying usability
-weights directly to CANO's per-feature allocation.]
-2.2 CANO Noise Allocation
-Given feature importance weights w_i and noise budget epsilon:
-    delta_i = epsilon * (w_i * n_features) * sign(z_i)
-where z_i ~ N(0,1) or gradient direction when available. The n_features scaling
-factor ensures equal total noise energy to baselines. The minimum weight floor
-(min_weight = 0.1) prevents attackers from exploiting negligibly-noised
-features.
-2.3 DQN Policy Training
-Adversarial co-evolution (Fig. 6): 50 simulated users, 20 samples/user,
-9 features.
-    State:   [feature_values, attack_confidence, privacy_budget, query_count]
-    Action:  per-feature noise allocation weights (softmax-normalized)
-    Reward:  alpha * privacy_gain - (1-alpha) * utility_cost
-Training alternates defender (CANO) and attacker (GradientBoosting retraining).
-2.4 Experimental Setup
-    Strategies:    Gaussian, FGSM, PGD, Carlini-Wagner, Laplace, CANO
-    Noise budgets: epsilon in {0.05, 0.10, 0.15, 0.20, 0.30, 0.50}
-    Attack models: Random Forest, Gradient Boosting, MLP
-    Datasets:      12 total -- 3 controlled synthetic (synth_small/medium/large),
-                   6 overlap/size-stress variants, 1 public keystroke
-                   (CMU 51-users), 1 real browser-fingerprint corpus (FP-Stalker,
-                   776 users), and cybersec_intrusion (2-user, OUT-OF-SCOPE)
-    Total configs: 68,885 (aggregate statistics over
-                   54,281 in-scope configs; utility
-                   metrics from the 3,529-row subset
-                   for which sparsity/KL/deviation/sensitivity were recorded).
-Scope note: cybersec_intrusion has 2 users (binary) -- not a valid k-class
-fingerprinting benchmark. It is retained in the per-dataset breakdown
-(Section 3.5) for completeness but excluded from all aggregate statistics,
-comparison tables, and significance tests.
-Data provenance. The 68,885 raw configurations span 19 evaluation runs from
-2026-03-22 through 2026-04-25. The paper quotes three different N values,
-which can otherwise look inconsistent:
-   N = 68,885   raw configurations across all 19 runs and all 12 datasets.
-   N = 54,281   in-scope configurations after excluding the 2-user
-                cybersec_intrusion dataset (the basis for Tables 1, 4, 5,
-                6 and Section 3 in general).
-   N =  3,529   utility-metric subset for which sparsity, KL divergence,
-                deviation, and sensitivity were recorded in-line (the
-                noise-quality fields were added to the evaluation script
-                in the 2026-04-05 run; earlier runs predate the
-                instrumentation and therefore do not contribute to
-                Table 3).
-Per-strategy row counts in Table 1 (range 8,460-10,423) differ because
-historical runs covered evolving subsets of the six strategies as the
-codebase matured (Laplace and PGD were added later than the original
-Gaussian/FGSM/CANO trio). All comparisons are strategy-paired within their
-own (dataset, attacker, epsilon, rep) tuples, so the unequal n's do not
-bias within-strategy means; they do, however, mean that the Table 1 totals
-should not be summed.
-================================================================================
-3. RESULTS
-================================================================================
-3.1 Overall Strategy Comparison (Adaptive Attack)  [See Figure 1, Figure 3]
-Table 1: Strategy comparison -- aggregate over in-scope datasets only.
-  Strategy       AccReduction       XferRed    NoiseL2   SNR(dB)        n
-  ------------------------------------------------------------------------
-  Gaussian       0.395 +/- 0.281   +0.411   0.595     9.7   10,423
-  Laplace        0.239 +/- 0.222   +0.391   0.556    11.2    8,460
-  FGSM           0.212 +/- 0.212   +0.472   0.642     9.8    9,641
-  PGD            0.129 +/- 0.171   +0.134   0.271    17.5    8,789
-  CANO (ours)    0.112 +/- 0.178   +0.263   0.435    15.5    8,508
-  C&W            0.001 +/- 0.011   -0.018   0.003    53.7    8,460
-Gaussian is the strongest adaptive-attack strategy. CANO outperforms only C&W
-on that metric. See Table 6 for significance.
-3.2 Transfer Attack Analysis  [See Figure 2]
-Table 2: Adaptive vs. transfer attack accuracy reduction (aggregate, in-scope).
-  Strategy       Adaptive  Transfer    Ratio      Gap
-  ------------------------------------------------------
-  CANO (ours)      0.112   +0.263     2.35x    +0.151
-  FGSM             0.212   +0.472     2.23x    +0.260
-  Laplace          0.239   +0.391     1.64x    +0.152
-  PGD              0.129   +0.134     1.04x    +0.005
-  Gaussian         0.395   +0.411     1.04x    +0.016
-  C&W              0.001   -0.018       n/a    -0.019
-(Transfer numbers are from the 2026-04-05 utility-metric subset; the new
-fpstalker block does not yet have transfer values computed -- backfill
-planned. Adaptive numbers are aggregate over all 54,281 in-scope rows.)
-CANO's 2.35x transfer-to-adaptive ratio reflects more
-model-agnostic perturbations. Gaussian provides little additional transfer
-protection. C&W is counterproductive (negative transfer reduction).
-3.3 Noise Utility Metrics
-Table 3: Per-strategy noise-quality metrics (n=3,529
-in-scope rows for which metrics were recorded; all CANO rows now report valid
-values after the evaluation-script fix).
-  FGSM           sparsity=1.000  KL=0.946  deviation=0.1824  sensitivity=-0.200  (n=649)
-  CANO (ours)    sparsity=1.000  KL=0.581  deviation=0.1690  sensitivity=-0.190  (n=540)
-  Gaussian       sparsity=1.000  KL=0.511  deviation=0.1423  sensitivity=+0.110  (n=720)
-  Laplace        sparsity=1.000  KL=0.440  deviation=0.1750  sensitivity=-0.185  (n=540)
-  PGD            sparsity=0.825  KL=0.156  deviation=0.0730  sensitivity=-0.071  (n=540)
-  C&W            sparsity=1.000  KL=0.008  deviation=0.0009  sensitivity=-0.027  (n=540)
-CANO's noise structure is distinguishable from Gaussian: similar KL divergence
-but different sensitivity signature (negative for CANO, positive for Gaussian),
-reflecting CANO's intentional concentration on importance-ranked (not
-variance-ranked) features.
-3.4 Epsilon Sensitivity  [See Figure 1]
-Table 4: Accuracy reduction by noise budget (aggregate, in-scope).
-   Epsilon   CANO    FGSM   Gaussian   Laplace   PGD    C&W
-   ----------------------------------------------------------
-   0.05     0.010  0.054    0.104    0.013  0.018  0.001
-   0.10     0.035  0.103    0.195    0.060  0.037  0.001
-   0.15     0.084  0.183    0.312    0.152  0.048  0.001
-   0.20     0.131  0.258    0.431    0.253  0.072  0.001
-   0.30     0.188  0.317    0.609    0.382  0.178  0.001
-   0.50     0.226  0.374    0.750    0.576  0.433  0.002
-CANO's ratio to Gaussian goes from 0.09 at epsilon=0.05 to
-0.30 at epsilon=0.50.
-3.5 Per-Dataset Analysis  [See Figure 5]
-Table 5: Mean accuracy reduction by dataset.
-  cybersec_intrusion       users=  2  CANO= 0.034  Gauss= 0.192  FGSM= 0.094  Lap= 0.122  PGD= 0.088  [OUT-OF-SCOPE: 2-user binary task; excluded from aggregates]
-  fpstalker (real)         users=776  CANO= 0.276  Gauss= 0.340  FGSM= 0.282  Lap= 0.291  PGD= 0.176
-  keystroke_cmu_51users    users= 51  CANO=   n/a  Gauss= 0.657  FGSM= 0.357  Lap=   n/a  PGD= 0.413
-  overlap_10u_50s          users= 10  CANO= 0.046  Gauss= 0.307  FGSM= 0.247  Lap=   n/a  PGD=   n/a
-  overlap_20u_30s          users= 20  CANO= 0.041  Gauss= 0.218  FGSM= 0.187  Lap=   n/a  PGD=   n/a
-  synth_10u_50s            users= 10  CANO= 0.001  Gauss= 0.270  FGSM= 0.057  Lap=   n/a  PGD=   n/a
-  synth_20u_50s            users= 20  CANO= 0.001  Gauss= 0.378  FGSM= 0.135  Lap=   n/a  PGD=   n/a
-  synth_50u_20s            users= 50  CANO= 0.003  Gauss= 0.514  FGSM= 0.356  Lap=   n/a  PGD=   n/a
-  synth_5u_30s             users=  5  CANO=-0.006  Gauss= 0.208  FGSM= 0.072  Lap=   n/a  PGD=   n/a
-  synth_large              users= 20  CANO= 0.135  Gauss= 0.416  FGSM= 0.234  Lap= 0.304  PGD= 0.152
-  synth_medium             users= 10  CANO= 0.082  Gauss= 0.306  FGSM= 0.153  Lap= 0.198  PGD= 0.090
-  synth_small              users=  5  CANO= 0.116  Gauss= 0.287  FGSM= 0.190  Lap= 0.215  PGD= 0.109
-cybersec_intrusion is shown for completeness only. fpstalker is the real
-browser-fingerprint corpus (Vastel et al. [10]) and is the closest dataset
-to deployment conditions; on it, CANO closes most of the gap to Gaussian
-(0.276 vs 0.340), in contrast to the wider gaps on small-synthetic datasets.
-3.6 Statistical Significance (aggregate, in-scope)  [See Figure 4]
-Table 6: CANO vs each baseline, Welch t-test and Cohen's d.
-  CANO vs C&W               d= +0.880   p < 0.001       ***
-  CANO vs FGSM              d= -0.510   p < 0.001       ***
-  CANO vs Gaussian          d= -1.202   p < 0.001       ***
-  CANO vs Laplace           d= -0.631   p < 0.001       ***
-  CANO vs PGD               d= -0.093   p < 0.001       ***
-3.7 Adversarial Training Results (DQN Policy)  [See Figure 6]
-DQN policy trained over 30 adversarial rounds with 50 users:
-    Baseline attack accuracy:   74.8%
-    Final attack accuracy:      20.8%
-    Accuracy reduction:         54.0 percentage points
-    Noise magnitude:            0.6061
-    DQN training steps:         31,500
-    Final Gini coefficient:     0.009 (near-uniform)
-Uniform allocation emerges as the game-theoretic equilibrium against adaptive
-adversaries.
-================================================================================
-4. DISCUSSION
-================================================================================
-4.1 Key Findings
-(1) Noise scaling (n_features multiplier) is the most impactful design choice.
-(2) Gaussian is the strongest adaptive-attacker defense (d = -1.20
-    vs CANO).
-(3) CANO achieves a 2.35x transfer/adaptive ratio.
-(4) RL equilibrium is uniform allocation (Gini = 0.009).
-(5) CANO uses less noise than Gaussian (L2 = 0.435 vs
-    0.595) with higher SNR (15.5
-    vs 9.7 dB).
-(6) On the real FP-Stalker corpus (776 users, 34 attributes), CANO closes most
-    of the gap to Gaussian (0.276 vs 0.340) -- a notably tighter result than
-    the typical small-synthetic gap (e.g., synth_50u_20s: CANO 0.003 vs
-    Gaussian 0.514). This suggests the importance-weighted allocation
-    generalizes better when feature importance reflects real attribute
-    redundancy rather than synthetic noise.
-4.2 Limitations
-  - One real browser-fingerprint dataset (FP-Stalker, 776 users) plus
-    synthetic/semi-synthetic plus CMU keystroke. Larger real-world fingerprint
-    corpora (HTillmann; BrFAST extended) are the next integration.
-  - 9 synthetic features with artificial importance concentration.
-  - cybersec_intrusion (2 users) excluded from aggregates.
-  - Utility metrics reported on 3,529-row subset
-    (most-recent run only). Backfill re-run planned for older configs.
-  - RL at 50 users; scaling TBD.
-4.3 Future Work
-  (1) Backfill transfer-attack and noise-utility metrics (sparsity, KL,
-      deviation, sensitivity) for the new fpstalker block; the older
-      synthetic-only utility-metric subset (n=3,529) does not yet include
-      fpstalker rows, so Tables 2 and 3 are still computed on that subset
-      while Tables 1, 4, 5, 6 use the full 54,281-row aggregate.
-  (2) Extend utility-metric coverage across all historical evaluation runs.
-  (3) Formal DP guarantees for CANO's allocation mechanism.
-  (4) Larger RL training (1,000+ users); online policy updates in deployment.
-  (5) Theoretical analysis of conditions under which feature-weighted noise
-      achieves higher transfer efficiency than uniform noise.
-================================================================================
-5. CONCLUSION
-================================================================================
-CANO does not match Gaussian in raw adaptive-attack accuracy reduction
-(0.112 vs 0.395), but achieves a 2.35x
-transfer-to-adaptive ratio -- better model-agnostic behavior than Gaussian
-(1.04x) in the transfer setting, which better reflects real-world
-deployment. On the real FP-Stalker browser-fingerprint corpus the
-adaptive-attack gap also narrows substantially (CANO 0.276 vs Gaussian 0.340),
-reinforcing the case that importance-weighted allocation generalizes better
-under realistic conditions than the small-synthetic aggregate suggests.
-Contributions:
-  (1) Noise scaling correction: equal total noise energy while redistributing
-      by importance.
-  (2) Transfer efficiency result: feature-importance weighting produces more
-      model-agnostic perturbations.
-  (3) RL equilibrium finding: uniform noise is the game-theoretic equilibrium
-      against adaptive adversaries (Gini = 0.009 after 30 rounds).
-  (4) Real-world validation: the 540-row complete block on FP-Stalker
-      (776 users, 13,674 fingerprints) shows the synthetic CANO/Gaussian gap
-      narrows substantially under realistic feature distributions.
-================================================================================
-FIGURES
-================================================================================
-Figure 1: Accuracy reduction vs. noise budget epsilon, per strategy.
-          File: results/figures/fig1_accuracy_reduction_vs_epsilon.png
-Figure 2: Privacy-utility Pareto front.
-          File: results/figures/fig2_pareto_front.png
-Figure 3: Per-strategy heatmap of accuracy reduction across datasets.
-          File: results/figures/fig3_strategy_heatmap.png
-Figure 4: Statistical significance of CANO vs each baseline.
-          File: results/figures/fig4_statistical_significance.png
-Figure 5: Per-dataset accuracy reduction bars.
-          File: results/figures/fig5_per_dataset.png
-Figure 6: DQN adversarial training progress (attacker accuracy vs round).
-          File: results/figures/rl_training_progress.png
-================================================================================
-REFERENCES
-================================================================================
-[1] Laperdrix, P. et al. "Browser Fingerprinting: A Survey." ACM CSUR, 2020.
-[2] Goodfellow, I. et al. "Explaining and Harnessing Adversarial Examples."
-    ICLR, 2015.
-[3] Madry, A. et al. "Towards Deep Learning Models Resistant to Adversarial
-    Attacks." ICLR, 2018.
-[4] Carlini, N. & Wagner, D. "Towards Evaluating the Robustness of Neural
-    Networks." IEEE S&P, 2017.
-[5] Dwork, C. et al. "The Algorithmic Foundations of Differential Privacy."
-    Foundations and Trends in TCS, 2014.
-[6] Mnih, V. et al. "Human-level Control through Deep Reinforcement Learning."
-    Nature, 2015.
-[7] Eckersley, P. "How Unique Is Your Web Browser?" PETS, 2010.
-[8] Andriamilanto, N. and Allard, T. "BrFAST: a Tool to Select Browser
-    Fingerprinting Attributes for Web Authentication According to a
-    Usability-Security Trade-off." Companion Proceedings of the Web
-    Conference 2021 (WWW '21 Companion), pp. 1-4, ACM, Ljubljana, Slovenia,
-    April 2021. DOI: 10.1145/3442442.3458610.
-    Source + data assets: github.com/tandriamil/BrFAST (MIT License).
-[9] Andriamilanto, N., Allard, T., and Le Guelvouit, G. "FPSelect:
-    Low-Cost Browser Fingerprints for Mitigating Dictionary Attacks
-    against Web Authentication Mechanisms." ACM CCS 2020.
-    DOI: 10.1145/3427228.3427297. arXiv:2010.06404.
-[10] Vastel, A., Laperdrix, P., Rudametkin, W., and Rouvoy, R.
-    "FP-STALKER: Tracking Browser Fingerprint Evolutions."
-    IEEE Symposium on Security and Privacy (S&P), pp. 728-741, 2018.
-    DOI: 10.1109/SP.2018.00008.
-    Raw dataset (~21,809 fingerprints, 40 attributes):
-    github.com/Spirals-Team/FPStalker.
-[11] Tillmann, H. "Browser Fingerprinting: 93% der Nutzer hinterlassen
-    eindeutige Spuren." Technical report, henning-tillmann.de, October
-    2013. Dataset redistributed as part of the BrFAST assets [8].
-================================================================================
-Generated: 2026-04-26 02:30:00
-Data source: 19 merged eval_*.jsonl files (68,885 raw configs from
-             overnight runs 2026-03-22 through 2026-04-25, including the
-             complete 540-row fpstalker block from the systemd-service
-             resume run 2026-04-19); excluding cybersec_intrusion from
-             aggregates (54,281 in-scope configs).
-================================================================================