Data-driven cyber attacks are more practical than the existing model-based ones. This technical note, from the attacker's standpoint, investigates how an adversary should design a sensorac-tuator coordinated attack policy to degrade system performance while undetected, without exact knowledge of system matrices. To guarantee the attack undetected, a notion of alpha-probability L-2 stealthiness is introduced. L-2-gain is employed to measure the attack's impacts while ensuring the stealth capabilities. As a result, the attack goal is formulated as a data-based L-2-gain composite optimization problem. Under this framework, a new multiobjective adaptive dynamic programming (ADP) method is proposed for launching the attack policy, which extends the traditional single-objective ADP technique. A numerical example is presented to demonstrate the effectiveness of the proposed attack scenarios.