If you mean this paper by Rogaway, "Evaluation of Some Blockcipher Modes of Operation" (PDF, relevant page is 120) then it very clearly documents everything.
- $q_{\text{dec}}$ is the total amount of decryption queries the adversary asks. If you only accept one decryption query before using different keys, this can be approximated by $1$.
- $\sigma$ is the total number of block cipher calls the mode would make on the adversary's sequence of queries. So if all adversarial queries (encryption and decryption ones) together require 4 block cipher calls to process, then $\sigma=4$. For 500 MByte of Data (in the single query) this would imply $\sigma\approx 2^{26}$
Also note that the bound stated in the paper (note the additional square on the $\sigma$) is $$\leq \frac{q_{\text{dec}}}{2^\tau}+\frac{\sigma^2}{2^b}$$
Plugging in the above values yields an advantage of $\leq 1/2^\tau+2^{52}/2^b$ which bounds the success probability of any adversary. Now random noise usually isn't exactly a smart adversary so the probability of a (successful) forgery from random noise is even much lower.
So let's clarify on what a "block cipher call" actually is. Modes of operation like CCM are built on block-ciphers which get modelled as functions $C=E_K(P)$ for some plaintext $P$, some key $K$ and the corresponding ciphertext $C$ where $P$ and $C$ are the same, fixed size. One computation of $C$ from $K,P$ or of $P$ from $K,C$ is called one block cipher call (because in the implementation you call the block cipher function for this). Now for CCM mode, which I will skip describing here, you end up with about 2 block cipher calls per data block (one for privacy, one for authenticity), yielding the aforementioned $2^{26}\approx 2\cdot 500\cdot 10^6/16$.
