Swarm Learning (SL) integrates federated learning and blockchain to support decentralized privacy-preserving learning for smart healthcare. However, the resilience of its learning and aggregation process against malicious data injection remains underexplored. This study presents a vulnerability analysis of SL against data-level poisoning, including label-flipping and GAN-generated adversarial sample injection. We conduct experiments across COVID-19 Radiography, ECG Heartbeat Categorization, and Credit Card Fraud Detection datasets under IID and non-IID client distributions. Results show that SL and FL exhibit dataset-dependent robustness under poisoning attacks. SL often preserves recall under attack, but this benefit can be accompanied by substantial precision degradation, especially in ECG-based physiological signal classification. To mitigate these effects, we evaluate the Trimmed-Mean aggregation defense. The results show that Trimmed-Mean can recover performance in several attack settings, although its effectiveness depends on assumptions about the attacker budget and the separability of malicious updates from benign non-IID updates. This study highlights that blockchain-based coordination alone is insufficient for secure SL deployment; robust learning-layer safeguards are required for safety-critical IoMT applications.