The technology and cyberspace sector is losing billions each year to hardware security threats. The incidents of usage of counterfeiting chips are doubling each year. The Electronic Resellers Association International (ERAI) reported that in the year 2011 more than 1300 counterfeits were reported. The incidents were double of what were reported in 2008. The report from Federal Contracts acknowledges the threats emanating from counterfeit chips and says it threatens the successful operations of US Weapon Systems. Meanwhile, electronic counterfeiting of chips continues to be a very profitable business on the dark web by crooked operatives. Physical Unclonable Functions (PUFs) are emerging as hardware security primitives to deal with security issues such as cloning, hacking, copying, and detection of Trojans. PUFs are one-way physical structures embedded in chips to generate a unique signature for each chip. The well-known silicon-based PUFs are Arbiter PUF (APUF) and Ring Oscillator PUF (ROPUF). The PUF uses timing delays caused by manufacturing process variations to generate challenge-response pairs (CRPs) unique to each chip. APUFs and ROPUFs are observed to be vulnerable to modeling attacks.In this research, a novel hybrid PUF is proposed which is a combination of both types of delay based PUFs, to generate strong cryptographic keys. The proposed design uses the CRPs of APUF and ROs of ROPUF to generate an n-bit response corresponding to an n-bit challenge, whereas primitive PUFs generate a 1-bit response for an n-bit challenge. The CRPs produced using the proposed PUF are unique and random and can be considered as cryptographic keys. The experimental results show that the uniqueness of APUF and ROPUF CRPs increase by 23% and 19%, respectively; when applied through the proposed scheme. The average passing rate for randomness is observed to be 97%.The CRPs generated from the delay based PUFs are tested against machine learning attacks. The machine learning attacks are carried out considering different scenarios where the adversary has access to 50%, 70%, 80%, and 90% of the CRPs. The models are trained for four different best-optimizing algorithms: Adagrad, Adadelta, SGD, and NAdam. The results show that even after training for the same number of epochs, the average accuracy for the proposed PUF model is 7% compared to 56% and 72% of APUF and ROPUF, respectively. The lower prediction accuracy of the proposed PUF shows that CRPs generated from the proposed scheme are far more immune to machine learning attacks when compared to other delay based PUFs.