Ransomware attacks are hazardous cyber-attacks that use cryptographic methods to hold victims' data until the ransom is paid. Zero-day ransomware attacks try to exploit new vulnerabilities and are considered a severe threat to existing security solutions and internet resources. In the case of zero-day attacks, training data is not available before the attack takes place. Therefore, we exploit Zero-shot Learning (ZSL) capabilities that can effectively deal with unseen classes compared to the traditional machine learning techniques. ZSL is a two-stage process comprising of: attribute learning (AL) and Inference Stage (IS). In this regard, this work presents a new Deep Contractive Autoencoder based Attribute Learning (DCAE-ZSL) technique as well as an IS method based on Heterogeneous Voting Ensemble (DCAE-ZSL-HVE). In the proposed DCAE-ZSL approach, Contractive Autoencoder (CAE) is employed to extract core features of known and unknown ransomware. The regularization term of CAE helps in penalizing the classifier's sensitivity against the small dissimilarities in the latent space. On the other hand, in case of the IS, four combination rules Global Majority (GM), Local Majority (LM), Cumulative vote-against based Global Majority (CVAGM), Cumulative vote-for based Global Majority (CVFGM) are utilized to find the final prediction. It is empirically shown that in comparison to conventional machine learning techniques, models trained on contractive embedding show reasonable performance against zeroday attacks. Furthermore, it is shown that the exploitation of these core features through the proposed voting based ensemble (DCAE-ZSL-HVE) has demonstrated significant improvement in detecting zero-day attacks (recall=0.95) and reducing False Negative (FN= 6).