In cyber-physical systems (CPSs), it is typical that a sensor observes a dynamical process and transmits the state estimate to a remote estimator wirelessly. Security risks arise when a denial-of-service (DoS) attacker generates extra noise at some power level to reduce the successful transmission rate. Investigating the capability of such an attacker to endanger the system is an important research line in CPS security. However, most previous works have two restrictions, one is that the attacker has complete knowledge of the system, which is usually difficult to achieve, and the other is that the attack power level set is small and discrete, which reduces the attack effectiveness and is hard to be implemented in multi-process systems due to the curse of dimensionality. In this paper, we tackle these restrictions by establishing a continuous attack power design for a DoS attacker with limited information. We propose deep deterministic policy gradient (DDPG)-based attack designs in single-process and multi-process systems, respectively. Numerical simulations illustrate the advantages of DDPG-based attack designs over heuristic baselines and existing learning methods.