Slide
Centre Interdisciplinaire
de Recherche et d’Innovation
en Cybersécurité et Société
de Recherche et d’Innovation
en Cybersécurité et Société
1.
Valem, L. P.; Pedronette, D. C. G.; Allili, M. S.
Contrastive Loss Based on Contextual Similarity for Image Classification Article d'actes
Dans: G., Bebis; V., Patel; J., Gu; J., Panetta; Y., Gingold; K., Johnsen; M.S., Arefin; S., Dutta; A., Biswas (Ed.): Lect. Notes Comput. Sci., p. 58–69, Springer Science and Business Media Deutschland GmbH, 2025, ISBN: 03029743 (ISSN); 978-303177391-4 (ISBN), (Journal Abbreviation: Lect. Notes Comput. Sci.).
Résumé | Liens | BibTeX | Étiquettes: Adversarial machine learning, Classification accuracy, Contrastive Learning, Cross entropy, Experimental evaluation, Federated learning, Image classification, Image comparison, Image embedding, Images classification, Model generalization, Model robustness, Neighborhood information, Self-supervised learning, Similarity measure
@inproceedings{valem_contrastive_2025,
title = {Contrastive Loss Based on Contextual Similarity for Image Classification},
author = {L. P. Valem and D. C. G. Pedronette and M. S. Allili},
editor = {Bebis G. and Patel V. and Gu J. and Panetta J. and Gingold Y. and Johnsen K. and Arefin M.S. and Dutta S. and Biswas A.},
url = {https://www.scopus.com/inward/record.uri?eid=2-s2.0-85218461565&doi=10.1007%2f978-3-031-77392-1_5&partnerID=40&md5=cf885303646c3b1a4f4eacb87d02a2b6},
doi = {10.1007/978-3-031-77392-1_5},
isbn = {03029743 (ISSN); 978-303177391-4 (ISBN)},
year = {2025},
date = {2025-01-01},
booktitle = {Lect. Notes Comput. Sci.},
volume = {15046 LNCS},
pages = {58–69},
publisher = {Springer Science and Business Media Deutschland GmbH},
abstract = {Contrastive learning has been extensively exploited in self-supervised and supervised learning due to its effectiveness in learning representations that distinguish between similar and dissimilar images. It offers a robust alternative to cross-entropy by yielding more semantically meaningful image embeddings. However, most contrastive losses rely on pairwise measures to assess the similarity between elements, ignoring more general neighborhood information that can be leveraged to enhance model robustness and generalization. In this paper, we propose the Contextual Contrastive Loss (CCL) to replace pairwise image comparison by introducing a new contextual similarity measure using neighboring elements. The CCL yields a more semantically meaningful image embedding ensuring better separability of classes in the latent space. Experimental evaluation on three datasets (Food101, MiniImageNet, and CIFAR-100) has shown that CCL yields superior results by achieving up to 10.76% relative gains in classification accuracy, particularly for fewer training epochs and limited training data. This demonstrates the potential of our approach, especially in resource-constrained scenarios. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.},
note = {Journal Abbreviation: Lect. Notes Comput. Sci.},
keywords = {Adversarial machine learning, Classification accuracy, Contrastive Learning, Cross entropy, Experimental evaluation, Federated learning, Image classification, Image comparison, Image embedding, Images classification, Model generalization, Model robustness, Neighborhood information, Self-supervised learning, Similarity measure},
pubstate = {published},
tppubtype = {inproceedings}
}
Contrastive learning has been extensively exploited in self-supervised and supervised learning due to its effectiveness in learning representations that distinguish between similar and dissimilar images. It offers a robust alternative to cross-entropy by yielding more semantically meaningful image embeddings. However, most contrastive losses rely on pairwise measures to assess the similarity between elements, ignoring more general neighborhood information that can be leveraged to enhance model robustness and generalization. In this paper, we propose the Contextual Contrastive Loss (CCL) to replace pairwise image comparison by introducing a new contextual similarity measure using neighboring elements. The CCL yields a more semantically meaningful image embedding ensuring better separability of classes in the latent space. Experimental evaluation on three datasets (Food101, MiniImageNet, and CIFAR-100) has shown that CCL yields superior results by achieving up to 10.76% relative gains in classification accuracy, particularly for fewer training epochs and limited training data. This demonstrates the potential of our approach, especially in resource-constrained scenarios. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
2.
Valem, L. P.; Pedronette, D. C. G.; Allili, M. S.
Contrastive Loss Based on Contextual Similarity for Image Classification Article d'actes
Dans: G., Bebis; V., Patel; J., Gu; J., Panetta; Y., Gingold; K., Johnsen; M.S., Arefin; S., Dutta; A., Biswas (Ed.): Lect. Notes Comput. Sci., p. 58–69, Springer Science and Business Media Deutschland GmbH, 2025, ISBN: 03029743 (ISSN); 978-303177391-4 (ISBN), (Journal Abbreviation: Lect. Notes Comput. Sci.).
Résumé | Liens | BibTeX | Étiquettes: Adversarial machine learning, Classification accuracy, Contrastive Learning, Cross entropy, Experimental evaluation, Federated learning, Image classification, Image comparison, Image embedding, Images classification, Model generalization, Model robustness, Neighborhood information, Self-supervised learning, Similarity measure
@inproceedings{valemContrastiveLossBased2025a,
title = {Contrastive Loss Based on Contextual Similarity for Image Classification},
author = {L. P. Valem and D. C. G. Pedronette and M. S. Allili},
editor = {Bebis G. and Patel V. and Gu J. and Panetta J. and Gingold Y. and Johnsen K. and Arefin M.S. and Dutta S. and Biswas A.},
url = {https://www.scopus.com/pages/publications/85218461565?origin=resultslist},
doi = {10.1007/978-3-031-77392-1_5},
isbn = {03029743 (ISSN); 978-303177391-4 (ISBN)},
year = {2025},
date = {2025-01-01},
booktitle = {Lect. Notes Comput. Sci.},
volume = {15046 LNCS},
pages = {58–69},
publisher = {Springer Science and Business Media Deutschland GmbH},
abstract = {Contrastive learning has been extensively exploited in self-supervised and supervised learning due to its effectiveness in learning representations that distinguish between similar and dissimilar images. It offers a robust alternative to cross-entropy by yielding more semantically meaningful image embeddings. However, most contrastive losses rely on pairwise measures to assess the similarity between elements, ignoring more general neighborhood information that can be leveraged to enhance model robustness and generalization. In this paper, we propose the Contextual Contrastive Loss (CCL) to replace pairwise image comparison by introducing a new contextual similarity measure using neighboring elements. The CCL yields a more semantically meaningful image embedding ensuring better separability of classes in the latent space. Experimental evaluation on three datasets (Food101, MiniImageNet, and CIFAR-100) has shown that CCL yields superior results by achieving up to 10.76% relative gains in classification accuracy, particularly for fewer training epochs and limited training data. This demonstrates the potential of our approach, especially in resource-constrained scenarios. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.},
note = {Journal Abbreviation: Lect. Notes Comput. Sci.},
keywords = {Adversarial machine learning, Classification accuracy, Contrastive Learning, Cross entropy, Experimental evaluation, Federated learning, Image classification, Image comparison, Image embedding, Images classification, Model generalization, Model robustness, Neighborhood information, Self-supervised learning, Similarity measure},
pubstate = {published},
tppubtype = {inproceedings}
}
Contrastive learning has been extensively exploited in self-supervised and supervised learning due to its effectiveness in learning representations that distinguish between similar and dissimilar images. It offers a robust alternative to cross-entropy by yielding more semantically meaningful image embeddings. However, most contrastive losses rely on pairwise measures to assess the similarity between elements, ignoring more general neighborhood information that can be leveraged to enhance model robustness and generalization. In this paper, we propose the Contextual Contrastive Loss (CCL) to replace pairwise image comparison by introducing a new contextual similarity measure using neighboring elements. The CCL yields a more semantically meaningful image embedding ensuring better separability of classes in the latent space. Experimental evaluation on three datasets (Food101, MiniImageNet, and CIFAR-100) has shown that CCL yields superior results by achieving up to 10.76% relative gains in classification accuracy, particularly for fewer training epochs and limited training data. This demonstrates the potential of our approach, especially in resource-constrained scenarios. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
3.
Kadi, A.; Selamnia, A.; Houda, Z. A. E.; Moudoud, H.; Brik, B.; Khoukhi, L.
An In-Depth Comparative Study of Quantum-Classical Encoding Methods for Network Intrusion Detection Article de journal
Dans: IEEE Open Journal of the Communications Society, vol. 6, p. 1129–1148, 2025, ISSN: 2644125X (ISSN).
Résumé | Liens | BibTeX | Étiquettes: Adversarial machine learning, Cyber attacks, Embeddings, Encoding methods, Encoding techniques, Encodings, Intrusion Detection, intrusion detection system, Intrusion Detection Systems, Machine-learning, Network embeddings, Network intrusion, Quantum cryptography, Quantum efficiency, Quantum electronics, Quantum machine learning, Quantum machines, Quantum-classical, Quantum-classical encoding, Zero-day attack
@article{kadiInDepthComparativeStudy2025,
title = {An In-Depth Comparative Study of Quantum-Classical Encoding Methods for Network Intrusion Detection},
author = {A. Kadi and A. Selamnia and Z. A. E. Houda and H. Moudoud and B. Brik and L. Khoukhi},
url = {https://www.scopus.com/pages/publications/85217024576?origin=resultslist},
doi = {10.1109/OJCOMS.2025.3537957},
issn = {2644125X (ISSN)},
year = {2025},
date = {2025-01-01},
journal = {IEEE Open Journal of the Communications Society},
volume = {6},
pages = {1129–1148},
publisher = {Institute of Electrical and Electronics Engineers Inc.},
abstract = {In today's rapidly evolving cyber landscape, the growing sophistication of attacks, including the rise of zero-day exploits, poses critical challenges for network intrusion detection. Traditional Intrusion Detection Systems (IDSs) often struggle with the complexity and high dimensionality of modern cyber threats. Quantum Machine Learning (QML) seamlessly integrates the computational power of quantum computing with the adaptability of machine learning, offering an innovative approach to solving intricate and high-dimensional challenges. A key factor in QML's performance is the method used to encode classical data into quantum states, as it defines how data is represented and processed in quantum circuits. QML offers promising advances for IDS, particularly through hybrid quantum-classical models. This study presents an in-depth comparative analysis of quantum-classical data encoding techniques for QML-based IDS. To the best of our knowledge, this is the first study to comprehensively evaluate the performance impact of different quantum encoding methods and provide a thorough evaluation of their impacts on the overall model performances. To achieve this, we first present a comprehensive evaluation of quantum and classical data encoding techniques, focusing on four key encoding techniques namely, Amplitude Embedding, Angle Embedding, Instantaneous Quantum Polynomial (IQP) Encoding, and Quantum Approximate Optimization Algorithm (QAOA) Embedding. Then, we develop a hybrid quantum-classical QML model to analyze how each encoding affects classification performance for malicious traffic. Finally, we conduct extensive experiments using two well-known, real-world network attack datasets to assess the accuracy and efficiency of each encoding approach. Our obtained results show notable differences in classification accuracy, underscoring the importance of encoding choice in optimizing QML-based IDS. This study aims to advance the application of quantum methodologies in network security by identifying effective encoding strategies for intrusion detection. © 2025 IEEE.},
keywords = {Adversarial machine learning, Cyber attacks, Embeddings, Encoding methods, Encoding techniques, Encodings, Intrusion Detection, intrusion detection system, Intrusion Detection Systems, Machine-learning, Network embeddings, Network intrusion, Quantum cryptography, Quantum efficiency, Quantum electronics, Quantum machine learning, Quantum machines, Quantum-classical, Quantum-classical encoding, Zero-day attack},
pubstate = {published},
tppubtype = {article}
}
In today's rapidly evolving cyber landscape, the growing sophistication of attacks, including the rise of zero-day exploits, poses critical challenges for network intrusion detection. Traditional Intrusion Detection Systems (IDSs) often struggle with the complexity and high dimensionality of modern cyber threats. Quantum Machine Learning (QML) seamlessly integrates the computational power of quantum computing with the adaptability of machine learning, offering an innovative approach to solving intricate and high-dimensional challenges. A key factor in QML's performance is the method used to encode classical data into quantum states, as it defines how data is represented and processed in quantum circuits. QML offers promising advances for IDS, particularly through hybrid quantum-classical models. This study presents an in-depth comparative analysis of quantum-classical data encoding techniques for QML-based IDS. To the best of our knowledge, this is the first study to comprehensively evaluate the performance impact of different quantum encoding methods and provide a thorough evaluation of their impacts on the overall model performances. To achieve this, we first present a comprehensive evaluation of quantum and classical data encoding techniques, focusing on four key encoding techniques namely, Amplitude Embedding, Angle Embedding, Instantaneous Quantum Polynomial (IQP) Encoding, and Quantum Approximate Optimization Algorithm (QAOA) Embedding. Then, we develop a hybrid quantum-classical QML model to analyze how each encoding affects classification performance for malicious traffic. Finally, we conduct extensive experiments using two well-known, real-world network attack datasets to assess the accuracy and efficiency of each encoding approach. Our obtained results show notable differences in classification accuracy, underscoring the importance of encoding choice in optimizing QML-based IDS. This study aims to advance the application of quantum methodologies in network security by identifying effective encoding strategies for intrusion detection. © 2025 IEEE.
4.
Houda, Z. Abou El; Moudoud, H.; Brik, B.
When Federated Learning Meets Knowledge Distillation to Secure Consumer Edge Network Article de journal
Dans: IEEE Transactions on Consumer Electronics, vol. 71, no 2, p. 7192–7200, 2025, ISSN: 00983063 (ISSN).
Résumé | Liens | BibTeX | Étiquettes: Adversarial machine learning, Collaborative modeling, Communication efficiency, Differential privacy, Distributed environments, Edge consumer network, EDGE Networks, Federated learning, Knowledge distillation, Model training, Secure multi-party computation, Security challenges, TinyML
@article{abouelhoudaWhenFederatedLearning2025,
title = {When Federated Learning Meets Knowledge Distillation to Secure Consumer Edge Network},
author = {Z. Abou El Houda and H. Moudoud and B. Brik},
url = {https://www.scopus.com/pages/publications/105002607319?origin=resultslist},
doi = {10.1109/TCE.2025.3559004},
issn = {00983063 (ISSN)},
year = {2025},
date = {2025-01-01},
journal = {IEEE Transactions on Consumer Electronics},
volume = {71},
number = {2},
pages = {7192–7200},
publisher = {Institute of Electrical and Electronics Engineers Inc.},
abstract = {Consumer networks face several security challenges due to the distributed nature of edge devices and the sensitive data they handle. Federated Learning (FL) presents a promising paradigm for collaborative model training in distributed environments. However, its implementation in edge consumer networks raises concerns about model heterogeneity, communication efficiency, and reverse engineering attacks. To address these issues, in this paper, we introduce SKDFL, a novel framework that leverages Knowledge Distillation (KD) and Secure Multi-Party Computation (SMPC) techniques to enhance communication efficiency while preserving data privacy in edge consumer networks. Through the use of KD, the distilled knowledge is transmitted between devices, significantly reducing communication overhead. Additionally, we incorporate lightweight encryption mechanisms to protect soft-labels from reverse engineering attacks using SMPC. We evaluate our proposed framework using two public datasets and demonstrate its efficiency in reducing communication costs, achieving up to a 92.4% reduction compared to conventional FL methods. Moreover, SKDFL achieves high performances in terms of accuracy and F1-score in both binary and multi-class classification while preserving the privacy of clients. Our obtained results show the potential of SKDFL to address the challenges of communication efficiency and data privacy in FL for edge consumer networks, paving the way for secure and efficient collaborative learning in consumer networks. © 1975-2011 IEEE.},
keywords = {Adversarial machine learning, Collaborative modeling, Communication efficiency, Differential privacy, Distributed environments, Edge consumer network, EDGE Networks, Federated learning, Knowledge distillation, Model training, Secure multi-party computation, Security challenges, TinyML},
pubstate = {published},
tppubtype = {article}
}
Consumer networks face several security challenges due to the distributed nature of edge devices and the sensitive data they handle. Federated Learning (FL) presents a promising paradigm for collaborative model training in distributed environments. However, its implementation in edge consumer networks raises concerns about model heterogeneity, communication efficiency, and reverse engineering attacks. To address these issues, in this paper, we introduce SKDFL, a novel framework that leverages Knowledge Distillation (KD) and Secure Multi-Party Computation (SMPC) techniques to enhance communication efficiency while preserving data privacy in edge consumer networks. Through the use of KD, the distilled knowledge is transmitted between devices, significantly reducing communication overhead. Additionally, we incorporate lightweight encryption mechanisms to protect soft-labels from reverse engineering attacks using SMPC. We evaluate our proposed framework using two public datasets and demonstrate its efficiency in reducing communication costs, achieving up to a 92.4% reduction compared to conventional FL methods. Moreover, SKDFL achieves high performances in terms of accuracy and F1-score in both binary and multi-class classification while preserving the privacy of clients. Our obtained results show the potential of SKDFL to address the challenges of communication efficiency and data privacy in FL for edge consumer networks, paving the way for secure and efficient collaborative learning in consumer networks. © 1975-2011 IEEE.
5.
Soltani, N.; Nejadshamsi, S.; Houda, Z. A. El; Khoury, R.; Costa, K. A. P.; Falk, T. H.; Avila, A. R.
Enhancing Network Intrusion Detection Systems: A Multi-Layer Ensemble Approach to Mitigate Adversarial Attacks Article d'actes
Dans: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern., p. 39–44, Institute of Electrical and Electronics Engineers Inc., 2025, ISBN: 1062922X (ISSN); 979-833153358-8 (ISBN), (Journal Abbreviation: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.).
Résumé | Liens | BibTeX | Étiquettes: Adversarial machine learning, Adversarial networks, Classification (of information), Computer crime, Fast gradient sign method, Fast Gradient Sign Method (FGSM), Generative adversarial network, Generative Adversarial Network (GAN), Generative adversarial networks, Generative model, Generative Models, Intrusion Detection, Intrusion-Detection, Learning algorithms, Learning systems, Machine-learning, Multi-layers, Network intrusion, Network intrusion detection systems, Network layers, Network security, Second layer, Stackings
@inproceedings{soltaniEnhancingNetworkIntrusion2025,
title = {Enhancing Network Intrusion Detection Systems: A Multi-Layer Ensemble Approach to Mitigate Adversarial Attacks},
author = {N. Soltani and S. Nejadshamsi and Z. A. El Houda and R. Khoury and K. A. P. Costa and T. H. Falk and A. R. Avila},
url = {https://www.scopus.com/pages/publications/105033159769?origin=resultslist},
doi = {10.1109/SMC58881.2025.11342479},
isbn = {1062922X (ISSN); 979-833153358-8 (ISBN)},
year = {2025},
date = {2025-01-01},
booktitle = {Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.},
pages = {39–44},
publisher = {Institute of Electrical and Electronics Engineers Inc.},
abstract = {Adversarial examples can represent a serious threat to machine learning (ML) algorithms. If used to manipulate the behaviour of ML-based Network Intrusion Detection Systems (NIDS), they can jeopardize network security. In this work, we aim to mitigate such risks by increasing the robustness of NIDS towards adversarial attacks. To that end, we explore two adversarial methods for generating malicious network traffic. The first method is based on Generative Adversarial Networks (GAN) and the second one is the Fast Gradient Sign Method (FGSM). The adversarial examples generated by these methods are then used to evaluate a novel multilayer defense mechanism, specifically designed to mitigate the vulnerability of ML-based NIDS. Our solution consists of one layer of stacking classifiers and a second layer based on an autoencoder. If the incoming network data are classified as benign by the first layer, the second layer is activated to ensure that the decision made by the stacking classifier is correct. We also incorporated adversarial training to further improve the robustness of our solution. Experiments on two datasets, namely UNSW-NB15 and NSL-KDD, demonstrate that the proposed approach increases resilience to adversarial attacks. © 2025 IEEE.},
note = {Journal Abbreviation: Conf. Proc. IEEE Int. Conf. Syst. Man Cybern.},
keywords = {Adversarial machine learning, Adversarial networks, Classification (of information), Computer crime, Fast gradient sign method, Fast Gradient Sign Method (FGSM), Generative adversarial network, Generative Adversarial Network (GAN), Generative adversarial networks, Generative model, Generative Models, Intrusion Detection, Intrusion-Detection, Learning algorithms, Learning systems, Machine-learning, Multi-layers, Network intrusion, Network intrusion detection systems, Network layers, Network security, Second layer, Stackings},
pubstate = {published},
tppubtype = {inproceedings}
}
Adversarial examples can represent a serious threat to machine learning (ML) algorithms. If used to manipulate the behaviour of ML-based Network Intrusion Detection Systems (NIDS), they can jeopardize network security. In this work, we aim to mitigate such risks by increasing the robustness of NIDS towards adversarial attacks. To that end, we explore two adversarial methods for generating malicious network traffic. The first method is based on Generative Adversarial Networks (GAN) and the second one is the Fast Gradient Sign Method (FGSM). The adversarial examples generated by these methods are then used to evaluate a novel multilayer defense mechanism, specifically designed to mitigate the vulnerability of ML-based NIDS. Our solution consists of one layer of stacking classifiers and a second layer based on an autoencoder. If the incoming network data are classified as benign by the first layer, the second layer is activated to ensure that the decision made by the stacking classifier is correct. We also incorporated adversarial training to further improve the robustness of our solution. Experiments on two datasets, namely UNSW-NB15 and NSL-KDD, demonstrate that the proposed approach increases resilience to adversarial attacks. © 2025 IEEE.
6.
Valem, L. P.; Pedronette, D. C. G.; Allili, M. S.
Contrastive Loss Based on Contextual Similarity for Image Classification Article d'actes
Dans: G., Bebis; V., Patel; J., Gu; J., Panetta; Y., Gingold; K., Johnsen; M.S., Arefin; S., Dutta; A., Biswas (Ed.): Lect. Notes Comput. Sci., p. 58–69, Springer Science and Business Media Deutschland GmbH, 2025, ISBN: 03029743 (ISSN); 978-303177391-4 (ISBN), (Journal Abbreviation: Lect. Notes Comput. Sci.).
Résumé | Liens | BibTeX | Étiquettes: Adversarial machine learning, Classification accuracy, Contrastive Learning, Cross entropy, Experimental evaluation, Federated learning, Image classification, Image comparison, Image embedding, Images classification, Model generalization, Model robustness, Neighborhood information, Self-supervised learning, Similarity measure
@inproceedings{valemContrastiveLossBased2025,
title = {Contrastive Loss Based on Contextual Similarity for Image Classification},
author = {L. P. Valem and D. C. G. Pedronette and M. S. Allili},
editor = {Bebis G. and Patel V. and Gu J. and Panetta J. and Gingold Y. and Johnsen K. and Arefin M.S. and Dutta S. and Biswas A.},
url = {https://www.scopus.com/inward/record.uri?eid=2-s2.0-85218461565&doi=10.1007%2f978-3-031-77392-1_5&partnerID=40&md5=cf885303646c3b1a4f4eacb87d02a2b6},
doi = {10.1007/978-3-031-77392-1_5},
isbn = {03029743 (ISSN); 978-303177391-4 (ISBN)},
year = {2025},
date = {2025-01-01},
booktitle = {Lect. Notes Comput. Sci.},
volume = {15046 LNCS},
pages = {58–69},
publisher = {Springer Science and Business Media Deutschland GmbH},
abstract = {Contrastive learning has been extensively exploited in self-supervised and supervised learning due to its effectiveness in learning representations that distinguish between similar and dissimilar images. It offers a robust alternative to cross-entropy by yielding more semantically meaningful image embeddings. However, most contrastive losses rely on pairwise measures to assess the similarity between elements, ignoring more general neighborhood information that can be leveraged to enhance model robustness and generalization. In this paper, we propose the Contextual Contrastive Loss (CCL) to replace pairwise image comparison by introducing a new contextual similarity measure using neighboring elements. The CCL yields a more semantically meaningful image embedding ensuring better separability of classes in the latent space. Experimental evaluation on three datasets (Food101, MiniImageNet, and CIFAR-100) has shown that CCL yields superior results by achieving up to 10.76% relative gains in classification accuracy, particularly for fewer training epochs and limited training data. This demonstrates the potential of our approach, especially in resource-constrained scenarios. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.},
note = {Journal Abbreviation: Lect. Notes Comput. Sci.},
keywords = {Adversarial machine learning, Classification accuracy, Contrastive Learning, Cross entropy, Experimental evaluation, Federated learning, Image classification, Image comparison, Image embedding, Images classification, Model generalization, Model robustness, Neighborhood information, Self-supervised learning, Similarity measure},
pubstate = {published},
tppubtype = {inproceedings}
}
Contrastive learning has been extensively exploited in self-supervised and supervised learning due to its effectiveness in learning representations that distinguish between similar and dissimilar images. It offers a robust alternative to cross-entropy by yielding more semantically meaningful image embeddings. However, most contrastive losses rely on pairwise measures to assess the similarity between elements, ignoring more general neighborhood information that can be leveraged to enhance model robustness and generalization. In this paper, we propose the Contextual Contrastive Loss (CCL) to replace pairwise image comparison by introducing a new contextual similarity measure using neighboring elements. The CCL yields a more semantically meaningful image embedding ensuring better separability of classes in the latent space. Experimental evaluation on three datasets (Food101, MiniImageNet, and CIFAR-100) has shown that CCL yields superior results by achieving up to 10.76% relative gains in classification accuracy, particularly for fewer training epochs and limited training data. This demonstrates the potential of our approach, especially in resource-constrained scenarios. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.