Face Tampering Identification Based On Lightweight Networ

Most recent works in face forgery detection put themselves into the pursuit of performance while they ignore the importance of computational complexity.The extensive computation make it hard to embed the heavy networks into edge devices,like smart phones.Hence,we believe that light-weight neural networks will be inevitably vital in the future research of face forensics.However,compared with heavy networks,light-weight neural networks often reaches a relatively low performance due to their limited capacity.To solve the above problem,we propose light-weight modules aiming to enhance feature representation with introducing trivial complexity.Finally,we reach a better trade-off between performance and computational complexity.Inspired by the previous works of face forgery detection,we introduce the long-range connection provided by the Non-Local method into the light-weight neural networks,that captures the relation between different manipulated areas which can not accomplished by the convolutions with its local receptive fields.Meanwhile,we propose a noval light-weight NonLocal module based on features’ group-wise clustering to lightweightly construct the longrange connection,aiming to reach a better trade-off between performance and complexity of our model.On the other hand,frequency information has a non-ignorable merit to describe the subtle manipulated contents in face forgery detection.Considering the importance of frequency and the lightweight property of our model,we propose a novel frequency dynamic convolution module via aggregating frequency attention mechanism into dynamic convolution to lightweightly enlarge the dynamic capacity of our models.It utilizes frequency information adaptively choose the kernel weights dedicated to the input image instead of the sharing one,therefore it enhances the performance of light-weight models in face forgery detection.