The problem of signal extraction from received data contaminated with strong chaotic interference has been investigated and a novel approach to this problem is proposed based on the geometry of chaotic interference. According to this approach, chaotic attractor is first reconstructed from the received data, then in the way of projecting chaotic interference and harmonic signals into the local tangent space and its transverse space of the manifold in which the attractor lies, the two signals are successfully separated. Excellent results were achieved from the computer simulation experiments, which verify the effectiveness of the approach.