Kaspa is an instant confirmation transaction sequencing layer. Transactions sent to miners can be included immediately in the ledger, which is structured as a blockDAG, to support asynchronous state updates. Kaspa is based on the PHANTOM protocol, a scalable generalization of Nakamoto Consensus. Its design is faithful to the principles Satoshi embedded into Bitcoin -- proof-of-work mining, UTXO-formed isolated state, deflationary monetary policy, no premine, and no central governance. Kaspa is unique in its ability to support high block rates while maintaining the level of security offered by proof-of-work environments. Kaspa currently operates with 1 block per second. Down the road, core developers and researchers will work on stretching the capability to the limits—think 10 or even 100 blocks per second (enabling a miner of 1% of the hashrate to mine on average 1 block every second!).
Kaspa is a proof-of-work cryptocurrency which implements the PHANTOM GHOSTDAG (a form of DAG) protocol. This means unlike other blockchains it does not orphan blocks created in parallel, rather allows them to coexist and orders them in consensus. GHOSTDAG is a generalization of Nakamoto consensus (Bitcoin) meaning it is theoretically as secure as Bitcoin as it makes no additional assumptions. One of the problems with the Nakamoto consensus, that Kaspa seeks to solve, is that due to the frequency of orphan blocks in order to revert a transaction, the attacker only need to create slightly less blocks than the honest network. So a 51% attack actually requires less than 51% of hashing power.
This doesn’t seem like a big deal, and indeed there is little difference between a 50.1% attacker and a 49.9% attacker. The problem is that when you try to upscale the throughput of the network (by either increasing the block rate, or the block size) you unavoidably increase the orphan rate, whereby decreasing the security of the network.
The security of any blockchain relies on the fact that the delay between blocks is larger by orders of magnitude than the time it takes the entire network to learn of a new block. Parallel blocks are orphaned, whereby they decrease the growth rate of the honest chain. Overcoming this throughput/security tradeoff is the main motivation behind the GHOSTDAG protocol.
Instead of Kaspa being a single chain like in many other cryptocurrencies Kaspa utilizes a DAG, meaning a block can point to many other blocks not just one. Double Spending is prevented by the manner in which the blocks are ordered. The DAG is made into a chain, traversed, and all transactions which do not contradict previous transactions are included. This way of ordering is the backbone of GHOSTDAG.
The GHOSTDAG ordering has a special property called the freeloading bound (which appears as Lemma 12 in the paper). It essentially means that an attacker which wishes to revert arbitrarily old blocks can not use honest blocks in a meaningful way. The amount of blocks they could freeload from is bound by a constant (namely 4k blocks). This means that any attack which seeks to change the ordering of arbitrarily old blocks will run very soon into a race where freeloading does not actually provide the attacker with any advantage.
A computationally inferior attacker can not revert arbitrarily old blocks, regardless of the ratio between the block delay and the block propagation time.
All of this is covered in more details in the GHOSTDAG whitepaper
Other DAG based projects like Nano or IOTA are not generalizations of the Nakamoto consensus, meaning they cannot be proven to theoretically be as secure as Bitcoin. The level of decentralization is also questionable in other DAG based projects. Many of them are also proof of stake, which has its own problems. Kaspa is provably secure and arguably will be faster than those coins.
Here's the point view of one of Kaspa theoreticians, Shai Wyborski.