This benchmark finds no consistent overall benefit from Tiny Feature Enhancement (TFE) for floating debris detection relative to its added computation. We compare YOLO11 P2 variants with ADown, TFE, TinySCA, and RepC3, alongside other detectors, using complete training annotations and five seeds. The original dataset supports 150 main runs, 75 grouped cross-validation runs, and ten batch-sensitivity runs. On its 40-image test set, scratch-trained P2–ADown reaches 12.15±0.57% AP50–95 versus 10.83±0.98% for YOLO11n, but no overall contrast survives the exploratory Welch/BH correction. Pretrained YOLO11n reaches 12.10%, and the cross-validation advantage of P2–ADown is only 0.19 points. Twenty additional scratch-trained runs on public FloW-Img test four architectures on 800 held-out images. P2–ADown–TFE without TinySCA reaches 34.53±0.62% AP50–95, compared with 33.27±0.70% for P2–ADown and 34.36±0.85% for P2. Thus, TFE helps relative to P2–ADown on FloW, but its mean is close to plain P2. Twenty zero-shot RiSID evaluations remain below 1% AP50–95. These results support a dataset-dependent assessment of accuracy and cost, not a general TFE advantage or readiness for autonomous monitoring.
Machine learning (ML)-based Intrusion Detection Systems (IDS) have become a key defence against advanced cyberattacks targeting Industrial Control Systems (ICS) in modern smart grids. Among these threats, Man-in-the-Middle (MITM) attacks are especially dangerous because they compromise the trust and privacy of communication between the control systems and other critical components within the interconnected smart grid. Although ML models have demonstrated a good ability to detect MITM attacks, they remain highly vulnerable to adversarial attacks, where small, carefully crafted changes can deceive the system and allow harmful traffic to go unnoticed. Even though adversarial machine learning is becoming increasingly important, its impact on ML-based MITM detection in smart grids remains inadequately explored. This paper examines the robustness of several widely used ML models against various adversarial attack methods using the RESLab smart grid dataset, an emulated testbed that produced network traffic containing MITM attack scenarios and is suitable for evaluating ML-based intrusion detection under adversarial conditions. To strengthen the models, we propose an adversarial training method that incorporates these hard-to-detect samples into the retraining process. Experiments show that adversarial attacks substantially degrade model performance. For example, fast-gradient sign method (FGSM)-perturbed MITM attacks can reduce the detection accuracy of the Random Forest (RF) classifier by 17.48 percentage points for binary classification and 14.86 percentage points for multiclass classification, demonstrating the vulnerability of conventional ML-based detection models. In contrast, when retrained with adversarial samples, the RF classifier achieves 81.59% detection accuracy for binary classification and 81.31% for multiclass classification, respectively, under fixed-source FGSM attacks, while maintaining strong performance on clean, unperturbed data. These results show that adversarial retraining greatly improves the robustness and reliability of ML-based MITM attack detection for protecting critical smart grid systems.
Span masking is usually justified informally: contiguous masks are said to force phrase-level rather than token-level inference. We quantify that argument and then test it. A boundary-screening theorem shows that the evidence a context carries about a masked set passes entirely through the block boundaries, so it is bounded by a quantity scaling with the number of blocks rather than of masked tokens. The downstream prediction this motivates, pre-registered with fixed decision rules, is refuted under the tested design. Across a sixteen-fold range of span length at fixed masking ratio, probe accuracy varies by 0.80 points, against 2.0 to 4.3 point changes the same design resolves from auxiliary objectives and training length. Mask geometry therefore bounds the available evidence without predicting linear-probe representation quality at this scale. Two further predictions hold and a third in part: the Gaussian closed form matches direct computation to 10−13 nats, a radial conservation law for cosine objectives holds in a trained transformer to within 3 percent, and an auxiliary masked-language term degrades semantic probes while barely moving lexical ones. The theorem is exact for finite-order Markov chains; its natural-language counterpart is a character-level diagnostic, and the encoders are 4.5 M parameters trained on 74 M tokens, at which a random-initialised baseline is already strong.
Deep learning forecasters of stock index direction often report high accuracy, but results are rarely checked under leakage-free protocols, against naive baselines, or across markets. We compared compact deep networks, classical learners, and naive rules for next-day direction over 3910 out-of-sample sessions (2022–2025), with SET100, the S&P 500, and CSI 300 as confirmatory markets and SET50, whose test years an earlier single-market version of this work had used, as exploratory. All models shared one point-in-time contract with expanding windows, purged boundaries, and train-only scaling; tuned and extended arms (GRU, TCN, DLinear, PatchTST-style, and MLP networks) selected settings on validation windows only, and the comparison was repeated with stationary features. No network showed a significant advantage over logistic regression or persistence in any confirmatory market, with original or stationary features, and equivalence within ±2 points was not established, so small advantages can be neither confirmed nor excluded. Cross-entropy networks mostly collapsed to all-up forecasts (AUC 0.43–0.55). A tuned random forest on CSI 300 (55.0% balanced accuracy) outperformed an independence benchmark in the Pesaran–Timmermann test after joint adjustment but showed no significant paired improvement over the competing baselines and lost significance with stationary features; it remains a candidate for prospective testing, not evidence of dependable skill. The study evaluates directional forecasting, not trading profitability. Single protocol errors raised the raw accuracy to 80%, and reporting 100 minus the mean absolute percentage error gave 99%. Code and predictions are openly available.
Updating a TinyML intrusion detector can change its decisions even when an authentic artifact is correctly installed: preprocessing, numerical representation, learned parameters, and the operating threshold jointly define the deployed function. We formalize this release as a pipeline contract and distinguish authenticated installation, decision agreement with a named reference, and detection quality. Signed pipeline packages and journaled release selection implement the contract on an ESP32-S3 and two ESP32-C3 boards. Evaluation combines chronological TON_IoT scenarios, a separate CICIoT2023 release experiment, host–device numerical checks, and installation experiments. Retaining an earlier threshold changed 15% of the floating-point MLP’s decisions. Export rules exact in real arithmetic still changed implemented decisions (9.3% for a threshold-folded logistic-regression export in one scenario), and targeted boundary inputs revealed host–device discrepancies absent on ordinary traffic. Alternative quantization improved the tested operating point while changing the reference integer release. Signed packages of 448–1588 bytes, up to 800 times smaller than the application image, reduced on-device installation processing about 40-fold on S3 and 14–16-fold on C3 relative to raw and compressed full-image delivery. During package installation, 185 of 188 rail-confirmed physical supply interruptions returned an expected release, with no unexpected or mixed release observed. These results support treating the complete pipeline as the unit of deployment and validating its decisions against an explicit reference.
