Class-aware Latent Attention Replay (CLAR): Class Incremental Hazard Diagnosis for Resilient AI in the Manufacturing Industrial Internet

Avranshu Mahanty; Ran Jin; Yingyan Zeng

cs.LGOct 29, 2026

Class-aware Latent Attention Replay (CLAR): Class Incremental Hazard Diagnosis for Resilient AI in the Manufacturing Industrial Internet

Authors: Avranshu Mahanty, Ran Jin, Yingyan Zeng

Organizations: Department of Mechanical, Materials, and Industrial Engineering, University of Cincinnati, Cincinnati, OH · Department of Industrial and Systems Engineering, Virginia Tech, Blacksburg, VA

Abstract

Artificial Intelligence (AI) systems provide critical computation services in the Manufacturing Industrial Internet (MII), and their resilience depends on timely diagnosis of hazards across the data, AI pipeline, and cyber-physical layers. During operation, new hazard types keep emerging with only a few samples each, hazards co-occur, and the diagnosis model runs on memory-constrained computation nodes. Replay-based class-incremental learning (CIL) can learn new hazard classes while revisiting a small memory of past samples, but existing methods assume single-label, balanced streams, store high-dimensional raw samples, and select them independently of the classifier. We propose Class-aware Latent Attention Replay (CLAR), which combines a dual buffer that stores mostly compact latent representations with a few raw samples, an Attention Leverage Score (ALS) that upper-bounds the classifier logit and guides class-balanced reservoir sampling, and an exponential-moving-average teacher that slows latent-space drift. For an idealized buffer model, we prove that the buffer is class-balanced and follows an ALS-weighted class distribution, and that below a critical per-class capacity, ALS-guided replay has a lower replay-gradient error than class-balanced uniform replay. In the case study, On hazard diagnosis data from an MII testbed, CLAR achieves the highest mean average weighted F1 score among the evaluated replay methods with the same buffer size

Figures & tables

Figure 1: Overview of the incremental diagnosis model, MII testbed, and the CLAR framework.

Panel of Figure 2 — Figure 2: Overview of CLAR. After each computation task, the AI system in the MII yields the supervised learning performance $X_{i,t}^{(P)}$ and the runtime metrics $X_{i,t}^{(R)}$ . The diagnosis model encodes both modalities into a latent representation $z_{i}$ and scores each hazard class with class-specific attention parameters $(\mathbf{A}_{c},\mathbf{v}_{c})$ . The CLAR components (shaded) are the dual replay buffer, the ALS-guided class-balanced sampling, which uses $(\mathbf{A}_{c},\mathbf{v}_{c})$ to admit samples, and the EMA teacher, which anchors the latent space through $\mathcal{L}_{\mathrm{KD}}$ . Dashed arrows denote replay.

Class	Layer	Root cause factor and simulation
Class 0	Data	Sensor contamination or failure: % of contaminated or failed sensors among the ten inputs
Class 1	Data	Low signal-to-noise ratio (SNR): random noise injected at controlled scales
Class 2	Data	Distribution change of input variables: Gaussian mixture data with controlled KL divergence
Class 3	Data	Class imbalance: nonconforming-to-conforming ratio within a data batch
Class 4	AI pipeline	Pipeline singularity: % of pipelines forced to produce identical predictions

Table I: Hazard classes and the corresponding root cause factors [ 2 ] .

	$\rho_{\text{inc}}=10$			$\rho_{\text{inc}}=15$
Method	AvgF1 $\uparrow$	Forgetting $\downarrow$	BWT	AvgF1 $\uparrow$	Forgetting $\downarrow$	BWT
No Replay	$0.8932\pm 0.0194$	$0.0071\pm 0.0038$	$0.0468\pm 0.0118$	$0.8855\pm 0.0183$	$0.0131\pm 0.0129$	$0.0382\pm 0.0186$
iCaRL	$0.8871\pm 0.0144$	$\mathbf{0.0058\pm 0.0026}$	$0.0428\pm 0.0062$	$0.8891\pm 0.0060$	$\mathbf{0.0062\pm 0.0042}$	$0.0426\pm 0.0096$
ER	$0.9069\pm 0.0090$	$0.0191\pm 0.0125$	$0.0349\pm 0.0159$	$0.8937\pm 0.0200$	$0.0306\pm 0.0102$	$0.0254\pm 0.0153$
CBRS	$0.8831\pm 0.0316$	$0.0394\pm 0.0350$	$0.0145\pm 0.0318$	$0.8918\pm 0.0190$	$0.0269\pm 0.0096$	$0.0330\pm 0.0190$
FBCR	$0.9007\pm 0.0214$	$0.0152\pm 0.0129$	$0.0490\pm 0.0226$	$0.8926\pm 0.0230$	$0.0199\pm 0.0125$	$0.0394\pm 0.0227$

Table II: Performance with

\rho_{\text{base}}=10

and

\rho_{\text{inc}}\in\{10,15\}

(mean

\pm

std over 5 runs). Buffer capacity: 50 raw samples for the baselines; 49 raw samples (

\mathcal{B}_{x}

) and 150 latent representations (

\mathcal{B}_{z}

) for CLAR. The best value in each column is in bold.

Panel of Figure 3 — Figure 3: Weighted F1 score per task for $\rho_{\text{inc}}=10$ (left) and $\rho_{\text{inc}}=15$ (right).

Figure 4: Class 1 per-class F1 Score Comparison across 5 benchmark runs, showing the mean $\pm$ standard deviation for imbalance ratio of 10. Class 1 is introduced as a new class at task index $j=3$ .

Figure 5: Plot of parameters of the classifier illustrating the learned class-specific attention parameters for after the final task across classes $0$ through $4$ . The top row displays the trainable per-class attention vectors $\mathbf{v}_{c}\in\mathbb{R}^{d}$ (visualized as bar plots across feature dimensions), which scale feature activations independently. The bottom row shows heatmaps of the per-class attention matrices $\mathbf{A}_{c}\in\mathbb{R}^{d\times d}$ , capturing complex feature interaction geometries and cross-dimensional weightings for each hazard class.

Figure 6: Plot of ALS vs. Logit diagnostics. Gray points represent hazard-free samples, while red points denote hazardous samples for class $c$ . The decision boundary is established at $\text{logit}=0$ (corresponding to a classification threshold of $0.5$ )

Figure 7: Plot of Training samples used to train the incremental hazard diagnosis model. Illustrating representational consistency during incremental training (Algorithm 2 ). Current streaming training samples are marked in blue circles (Stream), batch-wise retrieved raw samples from $\mathcal{B}_{x}$ are shown as green diamonds, and retrieved latent representations from $\mathcal{B}_{z}$ are shown as orange triangles. The spatial overlap between the modalities highlights the consistency and alignment maintained by the dual-replay memory buffer across classes.

Replay-Based Continual Learning

Replay-based continual learning preserves prior-task knowledge by revisiting stored or generated experience while adapting models to new data.

See all 134 papers

Amnesia: A Stealthy Replay Attack on Continual Learning Dreams
Jun 10, 2026Ahmed Sharshar, Naveen Kumar Kummari, Mohsen GuizaniReplay-Based Continual Learning
Department of Computer Vision Mohamed bin Zayed University of Artificial Intelligence, AbuDhabi, UAE · Department of Machine Learning Mohamed bin Zayed University of Artificial Intelligence, AbuDhabi, UAE · Department of Machine Learning2026 Mohamed bin Zayed University of Artificial Intelligence, AbuDhabi, UAE
Label-free Industrial Fault Detection via Adversarial Inverse Reinforcement Learning: A System for Run-to-Failure Prognostics
Jul 25, 2026Dhiraj Neupane, Mohamed Reda Bouadjenek, Richard Dazeley +1Reinforcement Learning Industrial Anomaly Detection
Deakin University · Waurn Ponds, Victoria, Australia
Beyond Class Marginals: Bounding Rehearsal Gaps without Freezing Class Co-occurrence
Sep 22, 2026Congren Dai, Nat Roongjirarat, Fei YeClass-Incremental Learning Replay-Based Continual Learning
Imperial College London · King’s College London · University of Electronic Science and Technology of China

Jun 10, 2026J/K move · Enter open · S save

Amnesia: A Stealthy Replay Attack on Continual Learning Dreams

Ahmed Sharshar, Naveen Kumar Kummari, Mohsen Guizani

Department of Computer Vision Mohamed bin Zayed University of Artificial Intelligence, AbuDhabi, UAE · Department of Machine Learning Mohamed bin Zayed University of Artificial Intelligence, AbuDhabi, UAE · Department of Machine Learning2026 Mohamed bin Zayed University of Artificial Intelligence, AbuDhabi, UAE

Replay-Based Continual Learning

Open paper AskPDF arXiv

Continual learning (CL) models often use experience replay to reduce catastrophic forgetting, but their robustness to replay sampling interference remains underexplored. Existing CL attacks alter inputs or training pipelines (poisoning/backdoors) and rarely include explicit auditable constraints, limiting realism. Here, auditability means a monitor can verify compliance from sampler-visible telemetry - e.g., logged replay index/label statistics - by checking that the realized replay class histogram stays close to a nominal baseline and that replay rate is unchanged per batch and/or over a rolling window. We study a limited-privilege insider who controls only replay index selection, not pixels, labels, or model parameters, while staying within auditable limits such as queue priorities. We introduce Amnesia, a replay composition attack that maximizes degradation under two budgets: a visibility budget delta bounding the TV/KL divergence from a nominal class histogram p0, and a mass budget f fixing the replay rate. Amnesia has two steps: (i) compute lightweight class utilities, such as EMA loss or confidence, to tilt p0 toward harmful classes; and (ii) project the tilt back into the delta-ball using efficient KL (exponential tilt) or TV (balanced mass redistribution) optimizers. A windowed scheduler enforces rolling audits. Across challenging CL benchmarks and strong replay baselines, Amnesia consistently lowers final accuracy (ACC) and worsens backward transfer (-BWT). The KL variant delivers high impact while remaining largely undetected under multiple audit schemes, including per-batch and rolling-window checks. The TV variant is more damaging but easier to detect, especially under tight per-class constraints. These results expose index-only replay control as a practical, auditable threat surface in CL systems and establish a principled impact-visibility trade-off.

Abstract

Replay-Based Continual Learning

Explore similar work

Amnesia: A Stealthy Replay Attack on Continual Learning Dreams

Label-free Industrial Fault Detection via Adversarial Inverse Reinforcement Learning: A System for Run-to-Failure Prognostics

Beyond Class Marginals: Bounding Rehearsal Gaps without Freezing Class Co-occurrence