We propose Falcon, an end-to-end 3-party protocol for efficient private\ntraining and inference of large machine learning models. Falcon presents four\nmain advantages - (i) It is highly expressive with support for high capacity\nnetworks such as VGG16 (ii) it supports batch normalization which is important\nfor training complex networks such as AlexNet (iii) Falcon guarantees security\nwith abort against malicious adversaries, assuming an honest majority (iv)\nLastly, Falcon presents new theoretical insights for protocol design that make\nit highly efficient and allow it to outperform existing secure deep learning\nsolutions. Compared to prior art for private inference, we are about 8x faster\nthan SecureNN (PETS'19) on average and comparable to ABY3 (CCS'18). We are\nabout 16-200x more communication efficient than either of these. For private\ntraining, we are about 6x faster than SecureNN, 4.4x faster than ABY3 and about\n2-60x more communication efficient. Our experiments in the WAN setting show\nthat over large networks and datasets, compute operations dominate the overall\nlatency of MPC, as opposed to the communication.\n