Linear cryptanalysis is still a key method to analyze the security of symmetric-key primitives, especially reduced-round block ciphers, in which structural weaknesses are more revealed. This paper examines linear approximation attacks on reduced-round variants of substitution permutation networks (SPNs) and Feistel ciphers, including finding high bias linear expressions in S-box layers and key mixing. The approach combines the linear hull analysis and the statistical hypothesis testing to accumulate the bias during several rounds, thus providing a way to recover keys. Experimental analysis shows that reduced-round designs have quantifiable biases up to 0.125, and achieve successful key recovery probabilities above 92.4 percent with much less plaintext-ciphertext pairs than designs using full rounds.