The development of learning-based detectors for massive multi-input multi-output (MIMO) systems has been hindered by the inherent complexities arising from the problem's high dimensionality. To enhance scalability, most previous studies have adopted model-driven methodologies that integrate deep neural networks (DNNs) within existing iterative detection frameworks. However, these methods often lack flexibility and involve substantial computational complexity. In this paper, we introduce ChannelNet, a purely data-driven learning-based massive MIMO detector that overcomes these limitations. ChannelNet exploits the inherent symmetry of MIMO systems by incorporating channel-embedded layers and antenna-wise shared feature processors. These modules maintain equivariance to antenna permutations and enable ChannelNet to scale efficiently to large numbers of antennas and high modulation orders with low computational complexity, specifically $\mathcal{O}(N_t N_r)$, where $N_t$ and $N_r$ denote the numbers of transmit and receive antennas, respectively. Theoretically, ChannelNet can approximate any continuous permutation-symmetric function and the optimal maximum likelihood detection (ML) function with arbitrary precision under any continuous channel distribution. Empirical evaluations demonstrate that ChannelNet consistently outperforms or matches state-of-the-art detectors across different numbers of antennas, modulation schemes, and channel distributions, all while significantly reducing computational overhead. This study highlights the potential of purely data-driven designs in advancing efficient and scalable detectors for massive MIMO systems.