Traditional OAC-FL schemes mitigate FL communication overhead via wireless channel signal superposition for accelerated aggregation but suffer from low spectrum efficiency and inadequate privacy—encrypting data only in transmission while relying on plaintext for server-side aggregation, thereby risking sensitive information leakage if the server is compromised or data intercepted. To address these limitations, this paper proposes a novel Federated Learning scheme integrating Sparse Code Multiple Access (SCMA) and lightweight Homomorphic Encryption (HE). Specifically, a parameter-aware dynamic SCMA codebook is designed: centered on client parameter update signals, it constructs low-correlation sparse codebooks, adapts sparsity based on the update amplitude of the previous round, and disables positions of long-term small-variation parameters to eliminate invalid transmissions. Additionally, a lightweight HE mechanism based on the Learning With Errors (LWE) problem is embedded, where local updates are encrypted prior to transmission and aggregation is performed at the ciphertext level to avoid privacy leakage. Experiments on the MNIST dataset demonstrate that the proposed scheme reduces communication overhead compared to vanilla FL while enhancing privacy, providing a practical solution for FL deployment in resource-constrained, privacy-sensitive scenarios such as edge computing.