Time-series human activity data obtained from sensor technologies facilitate various applications in industry and daily life, such as activity recognition, motion or fall detection, and health analysis. Recent research has shown that person re-identification and soft-biometric recognition are feasible from these activity recordings, leading to privacy breaches. Consequently, anonymising the subject characteristics found in the sensor recordings while retaining data utility is of interest. Here, we present an anonymisation framework using a conditioned autoencoder-based GAN that allows for three anonymisation strategies for time-series human activity data experimented on two complementary datasets. The framework was visually verified with experiments on motion capture data before being applied to inertial measurement data. This framework reduces re-identification to \(0.52\%\) while maintaining data utility for activity recognition tasks. Further, we present a form of anonymisation using identity transfer with the help of deep feature interpolation. The method achieves over \(96\%\) successful identity transfer with high data utility.