AutoDetect: Automated Multi-Domain Detection

• AutoDetect is a set of domain-specific frameworks that automate detection tasks, spanning plasma diagnostics, security anomaly detection, LLM weakness identification, and clinical screening.
• The systems use varied methodologies—from image thresholding and trajectory linking to autoencoder-based reconstruction error and LLM-mediated multi-agent assessment—to suit unique domain challenges.
• Benchmark results demonstrate practical improvements in speed, accuracy, and robustness, with measurable performance gains and deployment in real-world applications.

AutoDetect is a reused label in the arXiv literature rather than a single canonical algorithm. It denotes several domain-specific automated detection frameworks, including a dust-particle detector and tracker for RF plasma sheaths, an autoencoder-based detector for patch poisoning in military object detection, a three-agent framework for exposing weaknesses in LLMs, and a smartphone-based dental screening pipeline for children’s first permanent molars (Zayachuk et al., 2010, Liezenga et al., 3 Sep 2025, Cheng et al., 2024, Jiang et al., 2023). The shared theme is automation of expert inspection, but the underlying machinery ranges from image thresholding and trajectory linking to reconstruction-error modeling and LLM-mediated iterative assessment.

1. Principal referents of the name

The term appears in multiple, technically unrelated research programs. The main referents in the supplied literature are summarized below.

System	Domain	Core mechanism
AutoDetect (Zayachuk et al., 2010)	Dust particles in RF plasma sheath	Preprocessing, thresholding, local maxima, TSD-based tracking
AutoDetect (Liezenga et al., 3 Sep 2025)	Poisoning attacks on object detection	Convolutional autoencoder with slice-error thresholding
AutoDetect (Cheng et al., 2024)	LLM weakness discovery	Examiner, Questioner, and Assessor agents
AutoDetect (Jiang et al., 2023)	Dental screening from smartphone images	YOLOv5/YOLOX with overlapping tiling and NMS

The broader material also places these systems in a wider landscape of automatic detection, anomaly detection, model search, and elbow detection. That landscape includes EAutoDet for differentiable NAS in object detection, ODIN for video-drift detection and recovery, ADFilter for collider-event anomaly detection, a fully automatic trace-gas plume detector for imaging spectrometers, and UAED for automatic elbow detection in model selection (Wang et al., 2022, Suprem et al., 2020, Chekanov et al., 2024, Růžička et al., 5 May 2026, Morgado et al., 2023).

2. Dust-particle detection and tracking in RF plasma sheaths

In plasma diagnostics, AutoDetect denotes a pipeline for the automatic detection and tracking of large amounts of individual dust particles in a RF plasma sheath (Zayachuk et al., 2010). The method is designed for high-speed video, with the implementation blueprint giving an example of 500 fps at $15\ \mu\text{m}/\text{pix}$. A central design goal is operation under challenging illumination conditions and even without applying any external light source to facilitate detection.

The pipeline begins with frame preprocessing. Because dust appears as dark spots on a bright plasma background, gray levels are inverted so that dust becomes bright. Static artifacts are removed by subtracting an average fixed background $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$, and fluctuating background can then be suppressed either by a gliding-window mean,

$\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$

or by a 2D FFT band-pass with a high-pass cutoff at spatial scale $L_{\rm max}$ and low-pass at $L_{\rm noise}$. Segmentation is performed by thresholding. The method supports a fixed global threshold, an automatic global threshold $T_{\rm hr,g}=M-\tfrac12 S$, and a locally adaptive threshold based on the $7\times7$ neighborhood mean.

Detection proceeds by traversing foreground pixels in descending-intensity order and following steepest-brightness ascent in the 8-neighborhood until a local peak is reached. Local maxima closer than a particle radius $R_p$ are merged, with the brighter pixel retained. An optional sub-pixel refinement replaces the peak by the brightness-weighted centroid

$$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$

Feature extraction records position and brightness, optionally blob area or eccentricity, and rejects detections whose brightness falls below $I_{\rm noise}$ or whose area lies outside expected bounds.

Tracking is governed by a maximum inter-frame displacement $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$0 and an optional recovery time $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$1. Four strategies are described. TS-1 minimizes total squared displacement over one-to-one assignments; TS-2 is an oriented variant that processes particles in order of increasing local candidate count to prune the combinatorial search; TS-3 augments the cost with brightness change, $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$2; and TS-4 extrapolates each track’s last $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$3 positions and links to the nearest detection within $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$4, though it is not used by default because of noise sensitivity and extra cost.

Evaluation is unusually explicit. The benchmark modulates RF power with a small $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$5 sinusoidal perturbation of $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$6 on top of $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$7 DC, inducing a $B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$8 oscillation of sheath thickness and luminescence. The method records 1000 frames at 500 fps in a small ROI under the cathode and scores tracks using a correlation

$B_{\rm fix}(x,y)=\langle F_i(x,y)\rangle$9

where $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$0 is spatially averaged frame intensity and $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$1 is vertical particle position. Correct tracks satisfy $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$2, while $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$3 signals mistracking. The reported optimal configuration uses automatic global thresholding, $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$4 px, $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$5, and TS-1, yielding more than $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$6 of tracks with $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$7 (Zayachuk et al., 2010).

The same work supplements the physical benchmark with Monte Carlo validation in a $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$8 px box over $\tilde F_i = F_i-\tfrac12(F_{i-1}+F_{i+1}),$9 frames, with velocities drawn from a Tsallis distribution and optional collective oscillation in $L_{\rm max}$0. Over 1000 realizations, the pure Tsallis case produces 30% perfect tracks and 75% global accuracy, while Tsallis plus $L_{\rm max}$1-oscillation yields 50% perfect tracks and 85% global accuracy. Detection error is reported as $L_{\rm max}$2 px, approximately $L_{\rm max}$3, and empirical runtime on a Pentium IV is about 10 s for 1000 frames with $L_{\rm max}$4 under automatic thresholding.

3. Autoencoder-based poisoning detection for object detectors

In security research, AutoDetect is a “data-only,” model-agnostic anomaly detector for identifying patch-based poisoning attacks against object detection pipelines in the military domain (Liezenga et al., 3 Sep 2025). It is built around a simple 2D convolutional autoencoder and does not require access to the target detector or image labels. The only stated inputs are a small clean set of images from the target domain, or a proxy domain, and the suspect data.

The architecture operates on RGB images of size $L_{\rm max}$5, with $L_{\rm max}$6 pixels in the reported experiments. The encoder contains four Conv2D$L_{\rm max}$7BatchNorm$L_{\rm max}$8ReLU$L_{\rm max}$9MaxPool$L_{\rm noise}$0 blocks, with channels doubling as $L_{\rm noise}$1. After four poolings, the latent representation is a $L_{\rm noise}$2 tensor. The decoder is symmetric, using nearest-neighbor upsampling and Conv2D$L_{\rm noise}$3BatchNorm$L_{\rm noise}$4ReLU, ending in a 3-channel reconstruction with a final Sigmoid.

Detection is based on localized reconstruction error rather than image-level mean loss. For an image $L_{\rm noise}$5, the method computes per-pixel squared reconstruction error $L_{\rm noise}$6. A sliding $L_{\rm noise}$7 window, with baseline slice size $L_{\rm noise}$8 px and stride 1, produces slice means $L_{\rm noise}$9. Slice-error values from a clean validation set are fit to a normal distribution $T_{\rm hr,g}=M-\tfrac12 S$0. At test time, the maximum slice error $T_{\rm hr,g}=M-\tfrac12 S$1 is compared to the threshold $T_{\rm hr,g}=M-\tfrac12 S$2; equivalently, the tail probability $T_{\rm hr,g}=M-\tfrac12 S$3 is used, and the image is declared “poisoned” when $T_{\rm hr,g}=M-\tfrac12 S$4.

Training uses the MS COCO train split, approximately 80 k images normalized to $T_{\rm hr,g}=M-\tfrac12 S$5, with standard random flips and crop to $T_{\rm hr,g}=M-\tfrac12 S$6, 200 epochs, batch size 64, Adam with $T_{\rm hr,g}=M-\tfrac12 S$7, $T_{\rm hr,g}=M-\tfrac12 S$8, and learning rate $T_{\rm hr,g}=M-\tfrac12 S$9. The reference distribution is fit on a held-out subset of 1,000 MS COCO validation images. The evaluation also introduces MilCivVeh, a custom military/civilian vehicle dataset containing 1,456 images, 2,438 object instances, and three classes: civilian car, military truck, and military tank. Poisoning is implemented as a modified BadDet Global Misclassification Attack with a baseline $7\times7$0 HTBD patch, random patch location, and linear alpha blending with $7\times7$1; poisoned training labels are changed to the target class “military truck.”

Results are reported as AUROC on balanced clean/poisoned test sets. AutoDetect attains 0.965 on MS COCO, 0.941 on MilCivVeh, and 0.970 on VOC2007, exceeding PAD, PaDiM, and TRACE where measured (Liezenga et al., 3 Sep 2025). At threshold $7\times7$2, the method achieves approximately 95% accuracy in flagging poisoned images. The implementation is also lightweight: autoencoder weights occupy approximately 20 MB, fitting the slice-error distribution takes less than 30 seconds, and inference per image, including the forward pass and sliding-window pooling, is below 25 ms on an NVIDIA RTX 8000 GPU.

The paper is explicit about limitations. AutoDetect requires a manually verified clean subset, is sensitive to expected patch size, and performs best when slice size is approximately equal to patch size. Very small patches below 10 px are difficult to detect because they produce low slice means. The authors also note that the method assumes patch errors elevate reconstruction loss, so more subtle clean-label perturbations such as “Color Backdoor” may evade detection (Liezenga et al., 3 Sep 2025).

4. Weakness detection in LLMs

A third referent of AutoDetect is a unified framework for automated weakness detection in LLMs (Cheng et al., 2024). Here the name does not denote image analysis but a structured assessment process inspired by educational testing. The framework uses three LLM-powered agents: an Examiner that builds a hierarchical taxonomy of categories and knowledge points for a task, a Questioner that generates targeted items based on response history, and an Assessor that infers new knowledge points from low-scoring failures.

The workflow has two nested loops. In the outer “circulation,” Examiner, Questioner, and Assessor successively expand the weakness taxonomy and the question set. In the inner iterative search, the Questioner proposes increasingly difficult items for a selected knowledge point, conditioned on a history $7\times7$3 of question, model-response, and score tuples. Weaknesses are recorded when the judged score falls below a threshold. The implementation pseudocode in the supplied material makes this explicit: seed questions are generated for each knowledge point, responses from the target model are scored, new questions are proposed iteratively for up to $7\times7$4 rounds, and low-scoring cases are recycled by the Assessor into new hypothesized gaps.

Performance is measured by the Weakness Identification Success Rate,

$7\times7$5

with a GPT-4 judge assigning scores from 1 to 10 and treating scores $7\times7$6 as errors. The framework is evaluated on instruction-following, mathematics, and coding. Benchmarks include IFEval with 541 verifiable prompts, GSM8k, MATH, and HumanEval. Tested models include Llama 2-7B/13B/70B-Chat, Mistral-7B-Instruct, Llama 3-8B/70B-Instruct, GPT-3.5-turbo, Claude-3-Sonnet, Mistral Large, and GLM-4-Air.

The reported weakness discovery rates are substantial. The paper states that the framework attains an identification success rate exceeding 30% in prominent models such as ChatGPT and Claude (Cheng et al., 2024). The detailed tables similarly show nontrivial ISR across subdomains: GPT-3.5-turbo has an average ISR of 36.1%, while Claude-3 has 31.0%. The same targeted data are then used for fine-tuning. For Llama 2-7B, IFEval-p improves from 32.3 to 42.5, GSM8k from 18.9 to 25.9, MATH from 2.5 to 4.7, and HumanEval from 13.4 to 18.7. For Llama 2-13B, IFEval-p rises from 34.3 to 43.3 and HumanEval from 17.7 to 24.4. The abstract summarizes the broader finding as performance gains of over 10% across several benchmarks for models including the Llama series and Mistral-7b.

The framework also includes an explicit reproducibility stack. Agents are implemented with a lightweight Python wrapper around OpenAI’s chat-completion API using gpt-4-1106-preview; fine-tuning uses LoRA with AdamW, learning rate $7\times7$7, 10% warmup, linear decay, five epochs, Deepspeed ZeRO-2, batch size 4, and 8×A100-80 GB GPUs; inference uses vLLM with greedy decoding (Cheng et al., 2024). Reported limitations include a few unsolvable generated items despite 98% question reasonableness in a human evaluation of 150 cases, run-to-run ISR fluctuations of $7\times7$8–4%, and possible self-evaluation bias from relying on GPT-4 as judge.

5. Applied visual detection: dentistry and driving scenes

In clinical smartphone imaging, AutoDetect denotes a pipeline for detecting caries and pit-and-fissure sealing requirements in children’s first permanent molars (Jiang et al., 2023). The target is a single intraoral photograph of a child aged 7–9, from which the system locates each first permanent molar and classifies it as “Caries,” “PFS required,” or “No PFS required.” The implementation uses small one-stage detectors—YOLOv5s6 and YOLOX-s—on $7\times7$9 inputs, with training based on SGD, momentum $R_p$0, weight decay $R_p$1, base learning rate 0.01 with warm-up and cosine annealing, batch size 16 on 2×RTX2080 Ti, and 300 epochs.

The key engineering feature is a $R_p$2 overlapping tiling strategy with 50% overlap, introduced because first permanent molars occupy only a small fraction of many home-taken photos. A lightweight SVM using 1D color-histogram features filters tiles lacking teeth, with 77.3% accuracy and 90.5% recall on “has-tooth” versus “background.” Detection is run on both the full image and the retained tiles, and predictions are merged by class-aware non-maximum suppression at IoU threshold 0.45. The curated dataset contains 4,563 images and 5,077 annotated molars, drawn from an original pool of 16,023 smartphone oral photos. On the test set, YOLOX-s with tiling achieves 72.3 mAP@0.5 versus 71.2 without tiling, while YOLOv5s6 reaches 70.9 versus 67.9. The deployed system runs as a WeChat mini-applet backed by a Flask server on Tencent Cloud, with about 200 ms latency per image on a cloud GPU, excluding network overhead (Jiang et al., 2023).

A related applied-detection example in the supplied material is a transformer-based vehicle detector for difficult road conditions (Fahad et al., 25 Feb 2025). Although the paper is titled “Automatic Vehicle Detection using DETR,” its implementation notes describe an AutoDetect system built on Co-DETR. The baseline DETR formulation uses a transformer encoder-decoder with learned object queries and Hungarian set prediction, while Co-DETR augments the final decoder layer with $R_p$3 auxiliary heads trained under distinct label-assignment strategies. The reported configuration uses a pretrained Swin Transformer backbone at input size $R_p$4, AdamW with learning rate $R_p$5, weight decay $R_p$6, batch size per GPU 1, 20 epochs, and auxiliary head weights $R_p$7 with $R_p$8 and $R_p$9.

The driving dataset, BadODD, contains 9,825 images, 78,943 vehicle/object instances, and 13 classes across daytime and nighttime, urban, rural, highway, and expressway scenes from nine Bangladeshi districts. On the reported mAP@[.50:.95] benchmark at confidence threshold 0.4, Co-DETR reaches 0.372 at epoch 1, 0.418 at epoch 4, and 0.438 at epoch 9, compared with YOLOv8m values of 0.236, 0.255, and 0.295 at the same epochs. The paper also notes night-scene gains of about 12% absolute mAP over YOLOv8m and rural/highway gains of about 8–10%, while inference remains approximately 12 fps on an RTX 3090 because auxiliary heads are discarded at test time (Fahad et al., 25 Feb 2025).

6. Related automatic-detection paradigms and methodological context

The supplied literature situates AutoDetect within a larger family of automated detection systems whose formal names differ but whose design goals are similar: reducing manual review, operating under distribution shift, and coupling candidate generation with domain-specific scoring. ODIN is one such example for visual data analytics (Suprem et al., 2020). It uses a dual-adversarial autoencoder, DA-GAN, with an encoder $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$0, a decoder $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$1, and discriminators in latent and image space. Drift is detected by clustering latent codes, constructing density bands, and monitoring KL divergence as out-of-band data accumulate into a temporary cluster. When drift stabilizes, ODIN trains specialized models such as YOLO-SPECIALIZED or YOLO-LITE for the new concept. On BDD100K dashboard-camera videos, the system reports 6× higher throughput, 2× higher accuracy, and 6× smaller memory footprint than a baseline without automated drift detection and recovery.

Autoencoder-based anomaly scoring also appears in high-energy physics through ADFilter (Chekanov et al., 2024). Events are converted into $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$2 rapidity–mass matrices, flattened to 1×1287 vectors, and processed by a 1287→800→400→200→400→800→1287 dense autoencoder with Leaky ReLU activations and MSE loss. The anomaly score is the per-event reconstruction loss, and a threshold $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$3 is selected so that a fixed fraction, often 5%, of background events lie above the loss. The tool then recomputes invariant-mass and cross-section histograms after the cut. In the supplied examples, acceptance corrections move a $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$4 exclusion from 2.0 TeV to about 2.5 TeV and produce an exclusion at $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$5 TeV for a charged-Higgs scenario that was not excluded without anomaly-region selection (Chekanov et al., 2024).

Remote sensing contributes another fully automated detection pipeline for trace-gas plumes (Růžička et al., 5 May 2026). The system combines a gas-specific matched filter, a 5-model U-Net ensemble with MobileNetV3 encoder, and a physics-based spectroscopic fit $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$6. Candidate plume regions are produced by thresholding $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$7 at $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$8, enforcing a minimum region size of $$x_c=\frac{\sum_k I_k x_k}{\sum_k I_k},\qquad y_c=\frac{\sum_k I_k y_k}{\sum_k I_k}.$$9, and then verifying them by fitting measured transmittance inside and outside the region. High-confidence plumes satisfy $I_{\rm noise}$0, which empirically yields approximately 100% precision and about 45% recall on a 3,175-scene validation set. In operational “daily digest” mode, the system processes about 80–90 scenes per day, averages 0.51 alerts per day, and maintains a false-positive rate $I_{\rm noise}$1 at $I_{\rm noise}$2 (Růžička et al., 5 May 2026).

Two further examples extend the methodological context. EAutoDet performs efficient differentiable NAS over backbone and FPN search spaces for object detection, using kernel reusing and dynamic channel refinement; EAutoDet-s is reported to achieve 40.1 mAP with 120 FPS on COCO test-dev with a search cost of 1.4 GPU-days (Wang et al., 2022). UAED, by contrast, is not a detector of objects or anomalies but an automatic elbow detector for model selection. It minimizes

$I_{\rm noise}$3

runs in $I_{\rm noise}$4 time, and is positioned as a likelihood-free alternative to AIC, BIC, and HQIC in tasks such as clustering, regression order selection, and feature selection (Morgado et al., 2023).

A common misconception is that AutoDetect names a single benchmark, architecture, or software stack. The cited literature instead shows a heterogeneous nomenclature. This suggests that the most stable meaning of AutoDetect is operational rather than algorithmic: the automation of detection or diagnosis tasks that had previously depended on manual expert review, using whatever combination of preprocessing, statistical decision rules, learned representations, and domain constraints the application requires.