Darknets facilitate sophisticated cyber threats, but their encrypted, noisy traffic presents significant challenges, including multi-label complexity (e.g., overlapping threat categories and intricate label correlations), severe class imbalance, and dynamic traffic patterns. Traditional machine learning methods often fail to address these nuances effectively. This paper introduces a proposed hybrid LSTM-CNN model tailored for the Darknet2020 dataset, combining sequential (LSTM) and spatial (CNN) feature extraction to capture both temporal dependencies and hierarchical patterns, improving multi-label discrimination. The proposed model specifically addresses label correlation and class imbalance through dual-task learning and synthetic oversampling, enabling robust discrimination of co-occurring threats. The novel preprocessing pipeline mitigates noise and imbalance through adaptive filtering, while the dual-network architecture enhances robustness against label correlation. The model achieves 99% primary label accuracy and 86% secondary label accuracy, outperforming baselines by 11–17%, demonstrating its potential for real-world darknet threat detection.