Diversity-aware replay exemplars teach intrusion detection systems without forgetting

Every time a new kind of cyberattack appears on a network, the machine learning models guarding that network face an uncomfortable choice. They can either retrain from scratch on the full history of traffic, an expensive operation that many security operations centers cannot afford, or they can try to learn the new attack on the fly, an approach that risks wiping out everything they already knew. Researchers call this second failure mode catastrophic forgetting, and it has become one of the central obstacles to deploying artificial intelligence in real-world cybersecurity, where network traffic is fundamentally non-stationary and novel attack classes emerge continually. A new framework named DARE-IDS, short for Diversity-Aware Replay Exemplars for Incremental Intrusion Detection Systems, published in Cluster Computing, now offers a technically detailed answer to this problem, and its results suggest that the choice between learning new attacks and remembering old ones may finally be a false dilemma.

The work, led by Yacine Khaldi of the Mathematics Department at École Normale Supérieure de Ouargla in Algeria, together with Hanane Azzaoui and Mohamed Ben Bezziane of Kasdi Merbah University and Amir Benzaoui of University 20 August 1955 in Skikda, tackles a specific branch of continual learning known as Class-Incremental Learning, or CIL. In this paradigm, a model is confronted with a sequence of tasks, each introducing attack classes it has never seen before, and it must integrate the new knowledge while preserving its performance on previously learned classes. The difficulty is that, unlike humans, neural networks do not naturally compartmentalize knowledge. When gradients from the new task flow through the network, they reshape the very feature representations that encoded the old attacks, and accuracy on the old classes collapses. For an intrusion detection system, this collapse is not an academic inconvenience. A model that forgets how to recognize a denial-of-service flood while learning to spot a SQL injection attempt leaves the network exposed to an attack it was once able to catch.

The most widely used remedy in CIL is exemplar replay. The idea is elegantly simple: the model keeps a small memory buffer containing a handful of stored examples from each previously learned class, and during each incremental training step it replays these stored exemplars alongside the new data. The replayed examples act as anchors, pulling the network’s representations back toward the regions of feature space where the old classes live. But replay memory is strictly constrained. In a security appliance, storage and compute budgets are tight, and the buffer can hold only a tiny fraction of the training data. This constraint transforms exemplar selection from a minor implementation detail into a decisive design choice. Which specific examples, out of thousands of recorded network flows, should occupy the precious slots in the buffer? The answer determines how well the model’s memory of past attacks survives the continual arrival of new ones.

Conventional selection strategies answer this question in two familiar ways. The first is random selection, which simply picks exemplars uniformly at random from each class. The second is uncertainty-based selection, which favors examples near the model’s decision boundary, on the theory that these hard cases carry the most information. Both approaches, however, share a blind spot: they ignore the internal structure of each class. Network traffic for a single attack type is rarely homogeneous. A port-scan attack, for instance, may manifest as a wide spectrum of flow patterns depending on the tools used, the timing, and the network context. A buffer filled with random or boundary-focused exemplars may cover only a narrow slice of that spectrum, leaving large portions of the class distribution unrepresented in memory. When the model later relies on the buffer to reconstruct its knowledge, it reconstructs a distorted, incomplete picture of the attack, and performance on the old class degrades.

DARE-IDS answers this blind spot with a diversity-aware exemplar selection strategy. At its core, the framework selects exemplars that maximize feature diversity within each class. Rather than choosing examples that are representative in a narrow statistical sense, it chooses examples that are spread apart in the model’s learned feature space, ensuring that the buffer covers as much of each class’s distribution as possible under strict memory budgets. The intuition is that a diverse set of anchors constrains the network’s representations more completely: wherever the gradient from new training data tries to distort the feature space, a well-spread set of replayed examples pushes back from many directions at once. This is related in spirit to the core-set approach developed for active learning, where a carefully chosen subset of data can approximate the information content of the full dataset, but DARE-IDS adapts the principle to the specific demands of incremental intrusion detection, where class balance and coverage of rare attack variants matter enormously.

The exemplar selection strategy is only one component of the hybrid framework. DARE-IDS combines replay with knowledge distillation, a technique popularized by Hinton and colleagues in which the outputs of the previous model version serve as soft targets for the new one, preserving the old model’s functional behavior even on inputs that are not in the buffer. It also employs focal loss, a loss function originally introduced for dense object detection, which down-weights the contribution of easy, well-classified examples and concentrates training signal on the hard cases. This is particularly valuable in intrusion detection, where class imbalance is severe: benign traffic vastly outnumbers attack traffic, and within the attack classes themselves, some categories are extremely rare. Finally, the framework applies balanced fine-tuning, a rebalancing stage that corrects the classifier’s bias toward recently learned classes, a well-documented pathology in incremental learning where the last-seen classes dominate the decision layer.

The evaluation protocol is deliberately demanding. The researchers tested DARE-IDS on two benchmark datasets, UNSW-NB15 and CIC-IDS2017, under a five-step class-incremental protocol. This means the attack classes were split into five sequential tasks, and the model had to learn them one after another, with no access to the full training data at any point after the corresponding task had passed. The results are striking. DARE-IDS achieved a final accuracy of 88.2 percent on UNSW-NB15 and 96.8 percent on CIC-IDS2017, consistently outperforming both random and uncertainty-based exemplar selection baselines. Beyond raw accuracy, the framework yielded the highest weighted F1 score across both datasets, a metric that accounts for class imbalance and therefore reflects performance on the attack categories that matter most for security.

Perhaps the most important findings concern rare attack classes and the shape of the accuracy trajectory over the incremental steps. DARE-IDS preserved performance on rare attack classes better than the baselines, which is a critical property in cybersecurity, where the attacks that are hardest to collect data for are often the ones that cause the most damage when they strike. The framework also produced smoother accuracy trajectories with reduced catastrophic forgetting, meaning that as new tasks arrived, the model’s performance on old tasks declined less abruptly and recovered more reliably. In practical terms, this robustness translates into a security model whose behavior is predictable across deployment cycles, rather than one that oscillates between competence and amnesia as the threat landscape evolves.

The implications reach beyond the two datasets tested. The researchers have made their code publicly available on GitHub, and the datasets they analyzed are publicly accessible, which means the framework can be scrutinized, reproduced, and extended by the broader community. The work builds on a growing body of research applying continual learning to intrusion detection, including exemplar-based approaches, few-shot class-incremental methods, and federated learning frameworks, but its specific contribution, the diversity-aware selection of replay exemplars, addresses a gap that prior work had left open. As networks grow more complex, with Internet of Things devices, cloud infrastructure, and increasingly sophisticated adversaries, the ability of a detection system to learn continually without forgetting is shifting from a research aspiration to an operational requirement. DARE-IDS demonstrates that the design of the memory buffer, the humble set of stored examples that anchor a model’s past, may be the difference between a security system that adapts to the threat landscape and one that repeatedly relearns what it once knew.

For the cybersecurity industry, the message is clear: the next generation of intrusion detection systems will not be trained once and deployed forever. They will be living models, updated as new attack classes emerge, and their effectiveness will depend on how well they manage the tension between plasticity and stability. DARE-IDS shows that this tension can be managed with careful engineering, and that the diversity of remembered examples, not their quantity, is the key to a model that learns new attacks while refusing to forget the old ones.

Subject of Research: A hybrid class-incremental learning framework that combines diversity-aware exemplar replay, knowledge distillation, focal loss, and balanced fine-tuning to enable intrusion detection systems to learn new attack classes while preserving prior knowledge.

Article Title: DARE-IDS: diversity-aware replay exemplars for intrusion detection systems

Article References: Khaldi, Y., Azzaoui, H., Benzaoui, A., & Ben Bezziane, M. (2026). DARE-IDS: diversity-aware replay exemplars for intrusion detection systems. Cluster Computing, 29(15), Article 860. https://doi.org/10.1007/s10586-026-06696-4

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06696-4

Keywords: intrusion detection systems, class-incremental learning, catastrophic forgetting, exemplar replay, machine learning, cybersecurity, UNSW-NB15, CIC-IDS2017, DARE-IDS, diversity-aware, replay, exemplars