A powerful Bit-Flip Attack (BFA), based on Row Hammer Attack (RHA), can precisely flip the most vulnerable bits in the memory system (i.e., DRAM) to crash Convolutional Neural Networks (CNNs) run on Internet-of-Things (IoT) devices. However, it is very difficult to detect BFA since most devices are usually with limited computation capability and unaware of the security issue. Therefore, BFA becomes one of the most crucial threats to IoT networks. To this end, we design a novel defense system termed Resilient Dual-mode Defense System (RIDES) to encourage IoT devices with social relationships (i.e., Social IoT (SIoT)) to collaborate on BFA detection in an online manner. Subsequently, a new online problem, Online Computing Unit Assignment Problem (OMAR), is formulated to optimize the total inference rate for detecting BFA. To address OMAR's challenges, we present an online algorithm, Socially-aware Checker Assign-ment Algorithm (SCAN), to achieve the optimal competitive ratio. Extensive experiment and simulation results manifest that RIDES effectively detects BFA and in average, SCAN increases the total inference rate by 8%-515% and reduces the average overhead per image by 31%-55% compared with other solutions.