In this paper, we propose a task-oriented semantic communication system that conveys only task-related information in next-generation networks. This is especially important for efficient and intelligent data transmission in 6G-based Internet of Things (6G-IoT) scenarios where bandwidth constraints, latency requirements, and data privacy are critical. Existing methods struggle to completely separate task-related and irrelevant information, leading to privacy concerns and suboptimal performance. To address this, in this paper, we propose an information bottleneck-inspired method called Contrastive Learning and Adversarial Separation (CLAD), which effectively captures task-related features and removes task-irrelevant information using contrastive learning and adversarial separating. In addition, considering the lack of a reliable and reproducible method to quantify the minimality of encoded feature vectors, we introduce the Information Retention Index (IRI), which is used as a surrogate measure of mutual information between encoded features and inputs. IRI reflects how minimal and information-rich a representation is, making it highly relevant for privacy-preserving and bandwidth-efficient 6G-IoT systems. Through extensive experiments, we demonstrate that CLAD outperforms state-of-the-art baseline models in terms of semantic extraction, task performance, privacy, and IRI, demonstrating that it is a promising building block for responsible, efficient, and reliable 6G-IoT services.
