Modern cryptographic systems provide provable security guarantees as long as secret keys of the system remain confidential. However, if adversary learns some bits of information about the secret keys the security of the system can be breached. Side-channel attacks (like power analysis, timing analysis etc.) are one of the most effective tools employed by the adversaries to learn information pertaining to cryptographic secret keys. An adversary can also tamper with secret keys (say flip some bits) and observe the modified behavior of the cryptosystem, thereby leaking information about the secret keys. Dziembowski et al. (JACM 2018) defined the notion of non-malleable codes, a tool to protect memory against tampering. Non-malleable codes ensure that, when a codeword (generated by encoding an underlying message) is modified by some tampering function in a given tampering class, if the decoding of tampered codeword is incorrect then the decoded message is independent of the original message.

In this dissertation, we focus on improving different aspects of non-malleable codes. Specifically, (1) we extend the class of tampering functions and present explicit constructions as well as general frameworks for constructing non-malleable codes. While most prior work considered ``compartmentalized" tampering functions, which modify parts of the codeword independently, we consider classes of tampering functions which can tamper with the entire codeword but are restricted in computational complexity. The tampering classes studied in this work include complexity classes $\mathsf{NC}^0$, and $\mathsf{AC}^0$. Also, earlier works focused on constructing non-malleable codes from scratch for different tampering classes, in this work we present a general framework for constructing non-malleable codes based on average-case hard problems for specific tampering families, and we instantiate our framework for various tampering classes including $\mathsf{AC}^0$. (2) The locality of code is the number of codeword blocks required to be accessed in order to decode/update a single block in the underlying message. We improve efficiency and usability by studying the optimal locality of non-malleable codes. We show that locally decodable and updatable non-malleable codes cannot have constant locality. We also give a matching upper bound that improves the locality of previous constructions. (3) We investigate a stronger variant of non-malleable codes called continuous non-malleable codes, which are known to be impossible to construct without computational assumptions. We show that setup assumptions such as common reference string (CRS) are also necessary to construct this stronger primitive. We present construction of continuous non-malleable codes in CRS model from weaker computational assumptions than assumptions used in prior work.