Selected work

Representative Papers

MAGIC-SSCIL: Manifold Anchoring and Geometric Incremental Calibration for Semi-Supervised Class Incremental Learning

Aug 05, 2026

This work addresses catastrophic forgetting and unreliable pseudo-labeling in semi-supervised class-incremental learning without storing past samples by proposing a continual learning framework based on a frozen backbone network and learnable adapters. The method generates hallucinated features as substitutes for historical data through soft-weighted geometric calibration and introduces a geometric structure alignment objective to stabilize the topology of the feature space, effectively mitigating feature drift. It integrates graph-propagated labeling, Gaussian sampling, teacher-student head matching, and classifier prototype anchoring. Evaluated on CIFAR-100, CUB-200, and ImageNet-R, the approach significantly outperforms existing semi-supervised and supervised class-incremental methods using only 1%–10% labeled data, with particularly notable gains in fine-grained, low-label regimes.

0 citationsRead paper

Image Encryption Algorithm Based on Convolutional Neural Networks and Dynamic S-Box Generation

Jun 18, 2026

This work addresses the vulnerability of traditional image encryption schemes that rely on fixed S-boxes lacking input dependency, rendering them susceptible to linear and differential cryptanalysis. To overcome this limitation, the authors propose a novel hybrid approach integrating convolutional neural networks (CNNs) with classical cryptographic principles. Specifically, a pre-trained CNN extracts salient features from the input image to dynamically generate a personalized S-box for pixel substitution. This method represents the first implementation of content-adaptive S-box generation driven by the plaintext image itself, substantially enhancing the nonlinearity, uniqueness, and resilience of the encryption process against statistical and structural attacks. Experimental results demonstrate superior performance over conventional techniques across multiple security metrics, including information entropy, histogram uniformity, pixel correlation, NPCR, and UACI, thereby achieving both heightened security and greater flexibility.

0 citationsRead paper

Challenger at MultiPRIDE: Is It Hate Speech or Reclaimed?

May 31, 2026

This work addresses the challenge of distinguishing hate speech from reclaimed language used by marginalized communities on social media—a task complicated by strong context dependence, severe class imbalance, and substantial label noise. The authors propose a concise and interpretable approach that first generates dense semantic text embeddings, then employs Cleanlab to filter noisy labels, followed by classification using logistic regression and multilayer perceptrons. By integrating these components, the method effectively mitigates the adverse effects of extreme data imbalance and annotation errors while preserving model interpretability. Evaluated on the MultiPRIDE shared task, the approach demonstrates robust performance, achieving notably high macro-averaged F1 scores, precision, and recall, thereby offering a novel and effective strategy for identifying reclaimed language in online discourse.

0 citationsRead paper

Peacemaker at ATE-IT: Automatic term extraction from Italian text for waste management data using encoder model

May 31, 2026

This study addresses the challenges of scarce annotated data and difficulties in multi-word term extraction under domain shift in Italian waste management texts. To tackle these issues, the authors propose a low-cost, interpretable approach to automatic term extraction based on a fine-tuned encoder-based pretrained language model. The method achieves efficient term identification under limited computational resources and is evaluated using both type-level and token-level metrics. When applied to Task A of the ATE shared task (ATE-IT), the approach yields stable and balanced performance in precision, recall, and F1 score, comparable to that of other participating teams. These results demonstrate its effectiveness and offer a reliable, scalable solution for term extraction in low-resource scenarios.

0 citationsRead paper

Simorgh at SemEval-2026 task 7: Region-Aware Hybrid Retrieval for Low-Resource Cultural Reasoning in Multilingual Question Answering

May 26, 2026

This work addresses the insufficient cross-lingual stability of large language models when handling culturally grounded question answering in low-resource languages. To mitigate this issue, the authors propose a region-aware hybrid retrieval approach that combines BM25 lexical matching with dense semantic retrieval, augmented by a region-specific cultural weighting heuristic. This strategy constructs structured prompts for the quantized Qwen3-14B model, enabling deterministic answer selection based on logit scores. Evaluated on the BLEnD multilingual benchmark, the method significantly improves both relevance and cross-lingual consistency in cultural commonsense question answering. The results demonstrate the advantages of hybrid retrieval over purely parametric reasoning while also highlighting persistent performance gaps between high- and low-resource languages.

0 citationsRead paper

Recent publications

Latest Papers

MAGIC-SSCIL: Manifold Anchoring and Geometric Incremental Calibration for Semi-Supervised Class Incremental Learning

Aug 05, 2026

This work addresses catastrophic forgetting and unreliable pseudo-labeling in semi-supervised class-incremental learning without storing past samples by proposing a continual learning framework based on a frozen backbone network and learnable adapters. The method generates hallucinated features as substitutes for historical data through soft-weighted geometric calibration and introduces a geometric structure alignment objective to stabilize the topology of the feature space, effectively mitigating feature drift. It integrates graph-propagated labeling, Gaussian sampling, teacher-student head matching, and classifier prototype anchoring. Evaluated on CIFAR-100, CUB-200, and ImageNet-R, the approach significantly outperforms existing semi-supervised and supervised class-incremental methods using only 1%–10% labeled data, with particularly notable gains in fine-grained, low-label regimes.

0 citationsRead paper

Image Encryption Algorithm Based on Convolutional Neural Networks and Dynamic S-Box Generation

Jun 18, 2026

This work addresses the vulnerability of traditional image encryption schemes that rely on fixed S-boxes lacking input dependency, rendering them susceptible to linear and differential cryptanalysis. To overcome this limitation, the authors propose a novel hybrid approach integrating convolutional neural networks (CNNs) with classical cryptographic principles. Specifically, a pre-trained CNN extracts salient features from the input image to dynamically generate a personalized S-box for pixel substitution. This method represents the first implementation of content-adaptive S-box generation driven by the plaintext image itself, substantially enhancing the nonlinearity, uniqueness, and resilience of the encryption process against statistical and structural attacks. Experimental results demonstrate superior performance over conventional techniques across multiple security metrics, including information entropy, histogram uniformity, pixel correlation, NPCR, and UACI, thereby achieving both heightened security and greater flexibility.

0 citationsRead paper

Challenger at MultiPRIDE: Is It Hate Speech or Reclaimed?

May 31, 2026

This work addresses the challenge of distinguishing hate speech from reclaimed language used by marginalized communities on social media—a task complicated by strong context dependence, severe class imbalance, and substantial label noise. The authors propose a concise and interpretable approach that first generates dense semantic text embeddings, then employs Cleanlab to filter noisy labels, followed by classification using logistic regression and multilayer perceptrons. By integrating these components, the method effectively mitigates the adverse effects of extreme data imbalance and annotation errors while preserving model interpretability. Evaluated on the MultiPRIDE shared task, the approach demonstrates robust performance, achieving notably high macro-averaged F1 scores, precision, and recall, thereby offering a novel and effective strategy for identifying reclaimed language in online discourse.

0 citationsRead paper

Peacemaker at ATE-IT: Automatic term extraction from Italian text for waste management data using encoder model

May 31, 2026

This study addresses the challenges of scarce annotated data and difficulties in multi-word term extraction under domain shift in Italian waste management texts. To tackle these issues, the authors propose a low-cost, interpretable approach to automatic term extraction based on a fine-tuned encoder-based pretrained language model. The method achieves efficient term identification under limited computational resources and is evaluated using both type-level and token-level metrics. When applied to Task A of the ATE shared task (ATE-IT), the approach yields stable and balanced performance in precision, recall, and F1 score, comparable to that of other participating teams. These results demonstrate its effectiveness and offer a reliable, scalable solution for term extraction in low-resource scenarios.

0 citationsRead paper

Simorgh at SemEval-2026 task 7: Region-Aware Hybrid Retrieval for Low-Resource Cultural Reasoning in Multilingual Question Answering

May 26, 2026

This work addresses the insufficient cross-lingual stability of large language models when handling culturally grounded question answering in low-resource languages. To mitigate this issue, the authors propose a region-aware hybrid retrieval approach that combines BM25 lexical matching with dense semantic retrieval, augmented by a region-specific cultural weighting heuristic. This strategy constructs structured prompts for the quantized Qwen3-14B model, enabling deterministic answer selection based on logit scores. Evaluated on the BLEnD multilingual benchmark, the method significantly improves both relevance and cross-lingual consistency in cultural commonsense question answering. The results demonstrate the advantages of hybrid retrieval over purely parametric reasoning while also highlighting persistent performance gaps between high- and low-resource languages.

0 citationsRead paper